)]}'
{"specs/kilo/approved/cells-scheduling.rst":[{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":45,"context_line":""},{"line_number":46,"context_line":"When an API request is made to build an instance it expects to receive a"},{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_0efdc43b","line":48,"updated":"2014-12-17 14:16:10.000000000","message":"and here comes the first problem : request_spec is not persistent yet and we think to address this in Lxxx","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":45,"context_line":""},{"line_number":46,"context_line":"When an API request is made to build an instance it expects to receive a"},{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_5553d338","line":48,"in_reply_to":"3a961159_0efdc43b","updated":"2014-12-17 16:44:24.000000000","message":"I should clarify this more.  I don\u0027t intend that we persist anything that\u0027s not currently persisted.  Mainly what I want to protect against is an API restart that loses all knowledge of the instance.  Currently if we restart a nova service handling a build request we will lose the request_spec but the instance is still queryable and shown to be in building or in error.  I want to have parity with that, so if there\u0027s a restart the instance is still visible but it would currently lose the request_spec.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":45,"context_line":""},{"line_number":46,"context_line":"When an API request is made to build an instance it expects to receive a"},{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_dde0b398","line":48,"in_reply_to":"3a961159_5553d338","updated":"2014-12-18 08:27:29.000000000","message":"Oh OK, so please clarify what you want to persist. I guess the instance UUID but perhaps something more ?","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_ee8878c8","line":50,"updated":"2014-12-17 14:16:10.000000000","message":"why are you consider the scheduler response as not good enough for waiting it ?","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_dd8ef3cf","line":50,"in_reply_to":"3a961159_d0f6f1f8","updated":"2014-12-18 08:27:29.000000000","message":"Thanks for clarifying, so basically you say that the current model needs to be preserved. Why a simple RPC cast to the global scheduler can\u0027t be done here ? I don\u0027t see why the current Nova API could be a bit modified ?","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_d0f6f1f8","line":50,"in_reply_to":"3a961159_e4aed7e6","updated":"2014-12-17 16:44:24.000000000","message":"I completely agree.  But I don\u0027t think the user experience is acceptable if this is made synchronous.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_d05a7129","line":50,"in_reply_to":"3a961159_ee8878c8","updated":"2014-12-17 16:44:24.000000000","message":"Because the scheduler can take up to the messaging timeout to respond.  API requests should not be dependent on that, they should be as immediate as possible and then Nova can handle the rest asynchronously.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"3cbae366eca87a77b62114dfe82716a9da70c51e","unresolved":false,"context_lines":[{"line_number":47,"context_line":"response in a timely manner, and that an instance uuid has been assigned and"},{"line_number":48,"context_line":"all relevant details of the request stored in a persistent way.  This presents"},{"line_number":49,"context_line":"a few challenges to overcome.  The API request needs to return before we have"},{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_e4aed7e6","line":50,"in_reply_to":"3a961159_ee8878c8","updated":"2014-12-17 15:03:56.000000000","message":"Yes, this needs qualifying - it\u0027s much much easier if we can consider this a synchronous call.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_6e63a81a","line":53,"updated":"2014-12-17 14:16:10.000000000","message":"although request_spec is quite formalized now, filter_properties is not... So it would mean you would redo the wheel while we\u0027re working on objectify RequestSpec meanwhile... Maybe it would be cool to either find some workaround way to not persist the request, or wait for the objectification to be done","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_5dba036f","line":53,"in_reply_to":"3a961159_505041d4","updated":"2014-12-18 08:27:29.000000000","message":"Ack, as said above, please provide us a view of what you want to persist","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":50,"context_line":"made a call to the scheduler in order to be responsive.  But since we have not"},{"line_number":51,"context_line":"called the scheduler we can\u0027t place the instance info into a cells instance"},{"line_number":52,"context_line":"table.  So I\u0027m proposing a table at the api level to store the details of the"},{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_505041d4","line":53,"in_reply_to":"3a961159_6e63a81a","updated":"2014-12-17 16:44:24.000000000","message":"Right.  I don\u0027t want to persist the scheduling stuff right now.  I just want to persist enough of the instance so that it\u0027s queryable in the API and if the build request stalls or fails that can be stored somewhere.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_8129858b","line":56,"updated":"2014-12-17 14:16:10.000000000","message":"Eh, the current non-cell scheduler returns only host. Do you plan to keep the CellV1 scheduler or extend the non-cell Scheduler to provide a (cell, host) tuple ?\n\nI\u0027m in strong favor of the latter","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_9008e9da","line":56,"in_reply_to":"3a961159_8129858b","updated":"2014-12-17 16:44:24.000000000","message":"The latter.  What I envision is that the non-cell scheduler returns a (cell, host) tuple and the deployer can configure an API level scheduler to not return a host if desired.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":53,"context_line":"API request so that they can be persisted before the API responds."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_fd63af17","line":56,"in_reply_to":"3a961159_9008e9da","updated":"2014-12-18 08:27:29.000000000","message":"++\n\nPlease clarify it so","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""},{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_4e7c2cb6","line":58,"updated":"2014-12-17 14:16:10.000000000","message":"IMHO, conductor could still afford this. I don\u0027t like the idea to create a new service just for being a scheduler client...","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"3cbae366eca87a77b62114dfe82716a9da70c51e","unresolved":false,"context_lines":[{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""},{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_c1470d76","line":58,"in_reply_to":"3a961159_4e7c2cb6","updated":"2014-12-17 15:03:56.000000000","message":"+1 for no new service","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""},{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_f0f84dc6","line":58,"in_reply_to":"3a961159_c1470d76","updated":"2014-12-17 16:44:24.000000000","message":"I agree.  I just wanted to mention the possibility for discussion.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a cell, or (cell, host)"},{"line_number":56,"context_line":"tuple.  This could be done in a greenthread in the nova-api service, conductor"},{"line_number":57,"context_line":"could be modified to run at the api level as well for this, or a new service"},{"line_number":58,"context_line":"could be added similar to the current cells service."},{"line_number":59,"context_line":""},{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_7d77bfd4","line":58,"in_reply_to":"3a961159_f0f84dc6","updated":"2014-12-18 08:27:29.000000000","message":"So we could put this in the alternatives section","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"},{"line_number":62,"context_line":"a host, or that an admin has targeted a host, so there should be an optional"},{"line_number":63,"context_line":"host argument passed down.  If that host argument is not set then the conductor"},{"line_number":64,"context_line":"service handling the build process within the cell should query the scheduler"},{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_216291d8","line":63,"updated":"2014-12-17 14:16:10.000000000","message":"so if host set, we just send the request to the attached cell with no scheduling, ack","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":60,"context_line":"Then once a scheduling decision has been made the build request needs to be"},{"line_number":61,"context_line":"sent into the cell.  At that point it\u0027s possible that the scheduler has picked"},{"line_number":62,"context_line":"a host, or that an admin has targeted a host, so there should be an optional"},{"line_number":63,"context_line":"host argument passed down.  If that host argument is not set then the conductor"},{"line_number":64,"context_line":"service handling the build process within the cell should query the scheduler"},{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_3003b5b9","line":63,"in_reply_to":"3a961159_216291d8","updated":"2014-12-17 16:44:24.000000000","message":"Yep.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_01bf55fd","line":68,"updated":"2014-12-17 14:16:10.000000000","message":"and why not considering that all of this is just a Scheduler API response ? Same scheduler code can address all of these things, provided we add a new Scheduler method for providing a cell, it would be left to operators to see if they want to deploy either one nova-scheduler process for handling all of this, or dedicated schedulers per cell","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"3cbae366eca87a77b62114dfe82716a9da70c51e","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_e41fb707","line":68,"in_reply_to":"3a961159_01bf55fd","updated":"2014-12-17 15:03:56.000000000","message":"The way I imagined this working is:\n\n  each cell has a v1-style scheduler in it, which returns a list of candidate hosts.\n\n  If a deployer doesn\u0027t want to put a scheduler in each cell, then we imagine a null-scheduler which returns all hosts in the cell.\n\n  The top-level scheduler picks from a list of (cell,host) pairs.\n\nI feel that the solution here is a bit more complex than that, but (a) I probably haven\u0027t understood what alaski wrote completely and (b) there\u0027s probably domain complexity that I don\u0027t understand, too.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_f07bed26","line":68,"in_reply_to":"3a961159_d544231d","updated":"2014-12-17 16:44:24.000000000","message":"I can see this working either way.  A deployer can configure and deploy their schedulers to match their cells but all decisions are made through a single call in the API, or an API scheduler can optionally return (cell, None) or (cell, host) and (cell, None) would require calling a scheduler within a cell.\n\nI don\u0027t want to push a particular direction right now, I just recorded one possibility and hoped that someone might present an alternative like bauzas has.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":2033,"name":"Belmiro Moreira","email":"moreira.belmiro.email.lists@gmail.com","username":"moreira-belmiro-email-lists"},"change_message_id":"e0cd9b468826cda571fba3c94b0d3ff4a3d40311","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_ea6460b4","line":68,"in_reply_to":"3a961159_d544231d","updated":"2015-01-07 21:27:00.000000000","message":"I really like the vision for nova-scheduler extension to support cells V2. \nHowever, should we have nova-scheduler changes as a strong dependency for cells V2? \nThe existing \"cell scheduler\" maybe is a good enough approach until nova-scheduler is ready to replace it.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"9f63ea25b37b8d151b6bd6e8be82e9ed59b319e1","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_d544231d","line":68,"in_reply_to":"3a961159_e41fb707","updated":"2014-12-17 16:23:22.000000000","message":"Well, I think all of this is just a good opportunity for leveraging the existing nova-scheduler by providing a new interface for giving V2 cells, instead of reproducing the same model where there is specific scheduler for cells, which is not really maintained...\n\nThat also comes to a different direction that the Nova one, where the scheduler is being splitted to be enough agnostic","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":65,"context_line":"for a host placement.  The reason for this optionality is that some deployments"},{"line_number":66,"context_line":"may want a single scheduler that picks a host up front and other deployments"},{"line_number":67,"context_line":"might want a scheduler to pick a cell and then for schedulers within each cell"},{"line_number":68,"context_line":"to pick a host."},{"line_number":69,"context_line":""},{"line_number":70,"context_line":"Alternatives"},{"line_number":71,"context_line":"------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_bd584757","line":68,"in_reply_to":"3a961159_f07bed26","updated":"2014-12-18 08:27:29.000000000","message":"+1 to what you say, I was unclear but that\u0027s basically my thoughts : we have one API but possibly multiple schedulers\n\nSorry for being unclear","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"4d4bacabff2ad278dd2b4076e13520d42a98d0a0","unresolved":false,"context_lines":[{"line_number":77,"context_line":""},{"line_number":78,"context_line":"A new table would be added to the \u0027nova_api\u0027 database for storing build request"},{"line_number":79,"context_line":"details and an instance uuid.  This starts to look a little like storing a task"},{"line_number":80,"context_line":"to be completed."},{"line_number":81,"context_line":""},{"line_number":82,"context_line":"REST API impact"},{"line_number":83,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_e14a4926","line":80,"updated":"2014-12-17 14:16:10.000000000","message":"Long story you know...","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"368a55a1a1cdb8b2a393c042ef317468ef0e2926","unresolved":false,"context_lines":[{"line_number":77,"context_line":""},{"line_number":78,"context_line":"A new table would be added to the \u0027nova_api\u0027 database for storing build request"},{"line_number":79,"context_line":"details and an instance uuid.  This starts to look a little like storing a task"},{"line_number":80,"context_line":"to be completed."},{"line_number":81,"context_line":""},{"line_number":82,"context_line":"REST API impact"},{"line_number":83,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_fd31ef0b","line":80,"in_reply_to":"3a961159_154f8bbf","updated":"2014-12-18 08:27:29.000000000","message":"So, based on your replies, here you plan to persist the instance UUID and request_spec still ? I\u0027m a bit lost, please clarify.\n\nA model would be great IMHO :-)","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":77,"context_line":""},{"line_number":78,"context_line":"A new table would be added to the \u0027nova_api\u0027 database for storing build request"},{"line_number":79,"context_line":"details and an instance uuid.  This starts to look a little like storing a task"},{"line_number":80,"context_line":"to be completed."},{"line_number":81,"context_line":""},{"line_number":82,"context_line":"REST API impact"},{"line_number":83,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_154f8bbf","line":80,"in_reply_to":"3a961159_e14a4926","updated":"2014-12-17 16:44:24.000000000","message":"I am well aware :)","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"3cbae366eca87a77b62114dfe82716a9da70c51e","unresolved":false,"context_lines":[{"line_number":153,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":154,"context_line":""},{"line_number":155,"context_line":"Documentation will be written describing the flow of an instance build and how"},{"line_number":156,"context_line":"and where scheduling decisions are made."},{"line_number":157,"context_line":""},{"line_number":158,"context_line":""},{"line_number":159,"context_line":"References"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_24f1bfcd","line":156,"updated":"2014-12-17 15:03:56.000000000","message":"TBH we could do with a rough version of this now, to provide some concrete examples for discussion.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":153,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":154,"context_line":""},{"line_number":155,"context_line":"Documentation will be written describing the flow of an instance build and how"},{"line_number":156,"context_line":"and where scheduling decisions are made."},{"line_number":157,"context_line":""},{"line_number":158,"context_line":""},{"line_number":159,"context_line":"References"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_d5a2634a","line":156,"in_reply_to":"3a961159_24f1bfcd","updated":"2014-12-17 16:44:24.000000000","message":"That\u0027s what this spec is :)\n\nI\u0027m basically saying that whatever we decide here will get moved into a devref explanation.\n\nIf we want some concrete examples for discussion I can put them into the spec for review.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"5d4e9b5c3197fc9b16beb7c0b7972b707c3eb2ec","unresolved":false,"context_lines":[{"line_number":153,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":154,"context_line":""},{"line_number":155,"context_line":"Documentation will be written describing the flow of an instance build and how"},{"line_number":156,"context_line":"and where scheduling decisions are made."},{"line_number":157,"context_line":""},{"line_number":158,"context_line":""},{"line_number":159,"context_line":"References"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_d2d09d55","line":156,"in_reply_to":"3a961159_d5a2634a","updated":"2014-12-18 09:30:39.000000000","message":"Thanks Andrew. I see that developing in the conversation between you and Sylvain now, which is great.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"3cbae366eca87a77b62114dfe82716a9da70c51e","unresolved":false,"context_lines":[{"line_number":159,"context_line":"References"},{"line_number":160,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":161,"context_line":""},{"line_number":162,"context_line":"``https://etherpad.openstack.org/p/kilo-nova-cells``"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_44532b39","line":162,"updated":"2014-12-17 15:03:56.000000000","message":"and https://etherpad.openstack.org/p/nova-cells-scheduling-requirements","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a59a88c7fa895debb9708e543b05815aa0e9ef14","unresolved":false,"context_lines":[{"line_number":159,"context_line":"References"},{"line_number":160,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":161,"context_line":""},{"line_number":162,"context_line":"``https://etherpad.openstack.org/p/kilo-nova-cells``"}],"source_content_type":"text/x-rst","patch_set":1,"id":"3a961159_3593a7fb","line":162,"in_reply_to":"3a961159_44532b39","updated":"2014-12-17 16:44:24.000000000","message":"Will add that.","commit_id":"5cc4f4e251d4dcca492b697aa6d320486bc65f65"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"46958a2c6489eed3e97209e40d7e1063c17e11f8","unresolved":false,"context_lines":[{"line_number":52,"context_line":"about an instance build request until it has been scheduled and created within"},{"line_number":53,"context_line":"a cell."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a (cell, host) tuple."},{"line_number":56,"context_line":"This could be done in a greenthread in the nova-api service or conductor could"},{"line_number":57,"context_line":"be modified and run at the api level for this.  For clarity, this would be"},{"line_number":58,"context_line":"handled by the nova-scheduler and not an addition to the current cells"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_c317d552","line":55,"updated":"2015-01-14 17:20:09.000000000","message":"(cell, host, node) I guess, but thats a minor thing","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"b2c6c0660f87e4165478e5515ab2dab22d12f9d7","unresolved":false,"context_lines":[{"line_number":52,"context_line":"about an instance build request until it has been scheduled and created within"},{"line_number":53,"context_line":"a cell."},{"line_number":54,"context_line":""},{"line_number":55,"context_line":"The scheduler then needs to be queried in order to pick a (cell, host) tuple."},{"line_number":56,"context_line":"This could be done in a greenthread in the nova-api service or conductor could"},{"line_number":57,"context_line":"be modified and run at the api level for this.  For clarity, this would be"},{"line_number":58,"context_line":"handled by the nova-scheduler and not an addition to the current cells"}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a930d6b_43c19c2d","line":55,"in_reply_to":"3a961159_c317d552","updated":"2015-01-21 10:03:00.000000000","message":"Tbp, I would vote for a (cell, node) as the scheduler is only knowing about ComputeNodes, the host being only an attribute of the node. Thanks for the catch anyway.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"46958a2c6489eed3e97209e40d7e1063c17e11f8","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_a615773e","line":66,"updated":"2015-01-14 17:20:09.000000000","message":"Do we need this complexity? Could we not call the scheduler in the cell instead, and return the full result? I could be missing something.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":2033,"name":"Belmiro Moreira","email":"moreira.belmiro.email.lists@gmail.com","username":"moreira-belmiro-email-lists"},"change_message_id":"4d9b45f8de0831c1e60fca3e040de36752336f22","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a930d6b_071225b1","line":66,"in_reply_to":"1a930d6b_632bf8eb","updated":"2015-01-21 23:20:32.000000000","message":"That\u0027s a great idea!","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"3f4a02c41890438d47b971f5b06b61595948974b","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_5af8db49","line":66,"in_reply_to":"3a961159_2c178455","updated":"2015-01-16 12:09:07.000000000","message":"Hmm, thats a good point.\n\nI was thinking we would have a nova-cells-scheduler \u0027client\u0027 that hides the complexity of the extra hops until that is rolled into the scheduler proper, so we can keep the API-\u003escheduler-\u003ecompute flow as simple as possible, and unlikely to change.\n\nBasically using an adapter pattern while we wait for the scheduler to do what we want it to do.\n\nHaving said that, both work fine. I am not opposed to either approach, just I would lean towards the adapter approach myself.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"35cef8b8c39a82bdd93f11bba66e610ecee45205","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_b621a5ff","line":66,"in_reply_to":"3a961159_5af8db49","updated":"2015-01-19 14:47:07.000000000","message":"I just now realized you\u0027re suggesting calling the api scheduler and then the cell scheduler from the api level, rather than calling the cell scheduler from within the cell.  I\u0027ll have to think on that, but it seems reasonable.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"24c9416d8541887ba089d1f31cbdfca82ea6b657","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_2c178455","line":66,"in_reply_to":"3a961159_a615773e","updated":"2015-01-15 14:14:05.000000000","message":"It\u0027s really a matter of ordering how the work is done.  Long term I expect that there can be one scheduler endpoint at the api level which delegates work to cell level schedulers if necessary. But I believe there is a lot more involved in that than doing what I\u0027ve suggested here.  As well as scheduler work being currently focused on cleaning up interfaces which I expect means it will be a while before everything can be hidden behind a single scheduler API call.\n\nWhat I\u0027m suggesting is an intermediate step to help out in the short/medium term which doesn\u0027t need to wait on work happening elsewhere.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"b2c6c0660f87e4165478e5515ab2dab22d12f9d7","unresolved":false,"context_lines":[{"line_number":63,"context_line":"Then once the build request reaches the cell another scheduler call can be made"},{"line_number":64,"context_line":"to pick a host.  This adds a slight amount of complexity but will allow cellsv1"},{"line_number":65,"context_line":"users to migrate over to cellsv2 without needing to wait on large reworkings of"},{"line_number":66,"context_line":"the scheduler."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":""},{"line_number":69,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a930d6b_632bf8eb","line":66,"in_reply_to":"3a961159_b621a5ff","updated":"2015-01-21 10:03:00.000000000","message":"Yeah we are lucky enough to have a scheduler client now so we can hide the logic","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":9420,"name":"Matthew Gilliard","email":"matthew.gilliard@hp.com","username":"gilliard"},"change_message_id":"1d56dd5ff1168411756cfdb72da25b2e97cc71c0","unresolved":false,"context_lines":[{"line_number":71,"context_line":""},{"line_number":72,"context_line":"There are different variations of this proposal, such as using a two tiered"},{"line_number":73,"context_line":"scheduling system like cellsv1.  This spec is currently a forum to discuss"},{"line_number":74,"context_line":"this proposal against those alternatives."},{"line_number":75,"context_line":""},{"line_number":76,"context_line":"Data model impact"},{"line_number":77,"context_line":"-----------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_b6cb6f48","line":74,"updated":"2015-01-19 14:07:38.000000000","message":"I haven\u0027t seen much dissent from the proposed change recently. Are we ready to narrow down the suggestions and work on more specific details?","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"35cef8b8c39a82bdd93f11bba66e610ecee45205","unresolved":false,"context_lines":[{"line_number":71,"context_line":""},{"line_number":72,"context_line":"There are different variations of this proposal, such as using a two tiered"},{"line_number":73,"context_line":"scheduling system like cellsv1.  This spec is currently a forum to discuss"},{"line_number":74,"context_line":"this proposal against those alternatives."},{"line_number":75,"context_line":""},{"line_number":76,"context_line":"Data model impact"},{"line_number":77,"context_line":"-----------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_f6724d11","line":74,"in_reply_to":"3a961159_b6cb6f48","updated":"2015-01-19 14:47:07.000000000","message":"Yes.  The next step is to split this spec into multiple detailed specs on different aspects of this effort.  We can discuss details here if desired, but the output will ultimately be some other specs.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"46958a2c6489eed3e97209e40d7e1063c17e11f8","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_06c64b9c","line":96,"updated":"2015-01-14 17:20:09.000000000","message":"We need to cache the instances inside the API, I guess we could use the cache to store the request, but thats getting a bit deep into the details.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"24c9416d8541887ba089d1f31cbdfca82ea6b657","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_b203ca0b","line":96,"in_reply_to":"3a961159_06c64b9c","updated":"2015-01-15 14:14:05.000000000","message":"I\u0027m not sure exactly what you mean here.  Where are instances being cached inside the API?","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"ef382eb0366a6cbc853e61327ec4351df348d42e","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_3ac5cf3c","line":96,"in_reply_to":"3a961159_2ff76c5b","updated":"2015-01-16 12:00:40.000000000","message":"I was really mean, we spoke about the need to cache at the API layer to avoid the need to query multiple databases when listing instances.\n\nI just wondered if we could use that cache *before* the instance exists, but thats probably a bit dump.\n\nI do wonder if we should just store the API request in a generic \"task\" format that we could use for other APIs that need something similar.\n\nBut anyways, its a detail, we certainly need to store this info somewhere that show server and list servers can see.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"35cef8b8c39a82bdd93f11bba66e610ecee45205","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_96458971","line":96,"in_reply_to":"3a961159_3ac5cf3c","updated":"2015-01-19 14:47:07.000000000","message":"Gotcha.  There has been discussion about a potential cache for instances here, but at this point it hasn\u0027t been considered necessary so it has been put off for later optimization work.  I also think there will be a mismatch of the models for an instance cache and a boot request.\n\nA task format would be good, but I don\u0027t want to get caught up designing that in this effort as long as we can remain flexible for a future move towards that.  Since this table will be storing ephemeral data that only needs to last the lifetime of a single boot request I think we\u0027ll be relatively free to adapt it over time without needing to worry too much about migrations and versioning.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"9b1408af909520fd19336a58be0672b3c5b1ea6f","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_2ff76c5b","line":96,"in_reply_to":"3a961159_3b47b51e","updated":"2015-01-15 18:38:58.000000000","message":"That\u0027s only persistent for one request as I understand it.","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"b2c6c0660f87e4165478e5515ab2dab22d12f9d7","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a930d6b_63605836","line":96,"in_reply_to":"3a961159_96458971","updated":"2015-01-21 10:03:00.000000000","message":"I\u0027m still torn on this proposal. As it\u0027s unlikely to land by Kilo, I\u0027m in huge favor of being the less explicit possible and be pragmatic if the RequestSpec object series is merged meanwhile","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"},{"author":{"_account_id":6062,"name":"jichenjc","email":"jichenjc@cn.ibm.com","username":"jichenjc"},"change_message_id":"17cc566216db2803f58468a5b1d73affd3db9aae","unresolved":false,"context_lines":[{"line_number":93,"context_line":"      `availability_zone`,"},{"line_number":94,"context_line":"      `security_groups`,"},{"line_number":95,"context_line":"      ..."},{"line_number":96,"context_line":"    )"},{"line_number":97,"context_line":""},{"line_number":98,"context_line":"REST API impact"},{"line_number":99,"context_line":"---------------"}],"source_content_type":"text/x-rst","patch_set":2,"id":"3a961159_3b47b51e","line":96,"in_reply_to":"3a961159_b203ca0b","updated":"2015-01-15 15:31:38.000000000","message":"https://github.com/openstack/nova/blob/master/nova/api/openstack/wsgi.py#L123\n\nguess this is what is suggested?","commit_id":"6c62aa335c2ebb72eedd11d06cde906679440368"}],"specs/liberty/approved/cells-scheduling-interaction.rst":[{"author":{"_account_id":4393,"name":"Dan Smith","email":"dms@danplanet.com","username":"danms"},"change_message_id":"6f3f3dd66d670d58a5420c81b0dcca29ea3a232e","unresolved":false,"context_lines":[{"line_number":38,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":39,"context_line":""},{"line_number":40,"context_line":"The scheduler will be queried by the api so that it knows which cell to pass"},{"line_number":41,"context_line":"the build to.  Because the instance table exists in a cell and not in the api"},{"line_number":42,"context_line":"in order to create an instance we will first need to know which cell to create"},{"line_number":43,"context_line":"it in.  The scheduler will continue to return a (host, node) tuple and the api"},{"line_number":44,"context_line":"will need to look up which cell that is in.  So the current select_destinations"}],"source_content_type":"text/x-rst","patch_set":5,"id":"1aca2d91_a2be69fd","line":41,"range":{"start_line":41,"start_character":74,"end_line":41,"end_character":77},"updated":"2015-05-06 14:49:16.000000000","message":"I think you need a comma here -- it took me a couple reads of this sentence to get what you were saying.","commit_id":"bcdc56b128bb05397ef0c3082d583436f20678ca"},{"author":{"_account_id":4393,"name":"Dan Smith","email":"dms@danplanet.com","username":"danms"},"change_message_id":"6f3f3dd66d670d58a5420c81b0dcca29ea3a232e","unresolved":false,"context_lines":[{"line_number":44,"context_line":"will need to look up which cell that is in.  So the current select_destinations"},{"line_number":45,"context_line":"interface will not need to change.  Querying the scheduler will take place"},{"line_number":46,"context_line":"after the API has returned a response so some shifting of logic will need to"},{"line_number":47,"context_line":"occur and greenthreads will need to be spawned to run that code."},{"line_number":48,"context_line":""},{"line_number":49,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":50,"context_line":"compute-\u003econductor loop.  Adding rescheduling at a cell level will be a later"}],"source_content_type":"text/x-rst","patch_set":5,"id":"1aca2d91_627971fc","line":47,"updated":"2015-05-06 14:49:16.000000000","message":"When I first read this, I thought you were proposing that we query the scheduler once for the cell decision and then again for the host decision. However, I think you\u0027re proposing that we just query once for host, which implies a cell, and then create the actual instance record in that cell, right?\n\nI think that this means that we have written the request-to-build record in the api database before we return, and that we have some potential failure and recovery logic to implement in the API, right? If that thread that is trying to place the instance fails to get a response from select_destinations(), we\u0027ll need to retry some number of times, right?\n\nAlso, I think that the API nodes are mostly stateless now, which means people expect to be able to shoot them in the head at any time. This will be changing that such that if you scheduler is backed up, killing off some API nodes could drop some build requests on the floor. When we start those back up, do we try to recover that, or just leave the instance in building state?\n\nMaybe it would be better to run some conductors for the API whose sole purpose is to queue and manage builds on behalf of the API? That way we know that when we kill conductor nodes we might be losing something, and we can also try to do some smart shutdown logic for those where we mark the instance as failed before we exit if we\u0027re killed while waiting for a placement decision.","commit_id":"bcdc56b128bb05397ef0c3082d583436f20678ca"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"872bef247b7e8477946f1b5190399ae5d184ea56","unresolved":false,"context_lines":[{"line_number":44,"context_line":"will need to look up which cell that is in.  So the current select_destinations"},{"line_number":45,"context_line":"interface will not need to change.  Querying the scheduler will take place"},{"line_number":46,"context_line":"after the API has returned a response so some shifting of logic will need to"},{"line_number":47,"context_line":"occur and greenthreads will need to be spawned to run that code."},{"line_number":48,"context_line":""},{"line_number":49,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":50,"context_line":"compute-\u003econductor loop.  Adding rescheduling at a cell level will be a later"}],"source_content_type":"text/x-rst","patch_set":5,"id":"1aca2d91_78311eec","line":47,"in_reply_to":"1aca2d91_627971fc","updated":"2015-05-06 15:14:26.000000000","message":"Yeah I share Dan\u0027s concerns : if we consider a call from the API to the scheduler, how can we make sure it\u0027s asynchronous and doesn\u0027t hold the API response ? Using a conductor cast was really nice for it and I prefer it to a spawn_n() call, just because it\u0027s how it\u0027s already done now, and that works :-)\n\nCouldn\u0027t we just assume that n-cond would necessarly been shipped colocated with n-api (the parent), exactly like Dan said ?","commit_id":"bcdc56b128bb05397ef0c3082d583436f20678ca"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"872bef247b7e8477946f1b5190399ae5d184ea56","unresolved":false,"context_lines":[{"line_number":48,"context_line":""},{"line_number":49,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":50,"context_line":"compute-\u003econductor loop.  Adding rescheduling at a cell level will be a later"},{"line_number":51,"context_line":"effort."},{"line_number":52,"context_line":""},{"line_number":53,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":54,"context_line":"to account for that during its placement decisions, but that is outside the"}],"source_content_type":"text/x-rst","patch_set":5,"id":"1aca2d91_b83e96b6","line":51,"updated":"2015-05-06 15:14:26.000000000","message":"Then we could consider a local conductor, not the \u0027global\u0027 conductor I mentioned.","commit_id":"bcdc56b128bb05397ef0c3082d583436f20678ca"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"956cefa1d678a3f77f890abcb4d809585e220ffc","unresolved":false,"context_lines":[{"line_number":52,"context_line":""},{"line_number":53,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":54,"context_line":"compute-\u003econductor loop, using the conductors within the cell.  Adding"},{"line_number":55,"context_line":"rescheduling at a cell level will be a later effort."},{"line_number":56,"context_line":""},{"line_number":57,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":58,"context_line":"to account for that during its placement decisions, but that is outside the"}],"source_content_type":"text/x-rst","patch_set":8,"id":"3afb71cf_0f2b0a32","line":55,"updated":"2015-06-05 17:04:50.000000000","message":"Hmm, this looses something on cells v1 that already has that, but lets do this in small steps.\n\nThe inside a cell retry is interesting, because that retry is now scoped within a cell, oh, nasty, nasty.\n\nI think we should come back to this and iterate a bit, its hard to get the general solution right away.","commit_id":"288c1ce70b53010512a90c41a5811baf1fa518e0"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"0f7e584836e627b7b8da51ca76594c4ceb77cb71","unresolved":false,"context_lines":[{"line_number":52,"context_line":""},{"line_number":53,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":54,"context_line":"compute-\u003econductor loop, using the conductors within the cell.  Adding"},{"line_number":55,"context_line":"rescheduling at a cell level will be a later effort."},{"line_number":56,"context_line":""},{"line_number":57,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":58,"context_line":"to account for that during its placement decisions, but that is outside the"}],"source_content_type":"text/x-rst","patch_set":8,"id":"3afb71cf_b40321b8","line":55,"in_reply_to":"3afb71cf_0f2b0a32","updated":"2015-06-05 20:37:04.000000000","message":"Cellsv1 only reschedules at the cell level if a picked cell is down, i.e. the message queue doesn\u0027t accept the message.  We can keep that functionality.  What isn\u0027t covered is if scheduling fails within a cell it won\u0027t kick back up to pick another cell.  Something that doesn\u0027t exist in v1.\n\nThis will need to be rethought when we get to multiple cells, but for now we\u0027re just targeting a single cell case.  Really I could reword this to not even consider multiple cells yet.","commit_id":"288c1ce70b53010512a90c41a5811baf1fa518e0"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"d7a88f91d93f8ec6b08db34af02e074aabbdff02","unresolved":false,"context_lines":[{"line_number":52,"context_line":""},{"line_number":53,"context_line":"Rescheduling will still take place within a cell via the normal"},{"line_number":54,"context_line":"compute-\u003econductor loop, using the conductors within the cell.  Adding"},{"line_number":55,"context_line":"rescheduling at a cell level will be a later effort."},{"line_number":56,"context_line":""},{"line_number":57,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":58,"context_line":"to account for that during its placement decisions, but that is outside the"}],"source_content_type":"text/x-rst","patch_set":8,"id":"3afb71cf_727a8f25","line":55,"in_reply_to":"3afb71cf_b40321b8","updated":"2015-06-11 11:50:56.000000000","message":"So the bit (i think) we loose is if you can\u0027t space in any cells it waits a bit for the state to update, then tries again. But that feature doest really work right now anyways, so its no big loss.","commit_id":"288c1ce70b53010512a90c41a5811baf1fa518e0"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"956cefa1d678a3f77f890abcb4d809585e220ffc","unresolved":false,"context_lines":[{"line_number":56,"context_line":""},{"line_number":57,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":58,"context_line":"to account for that during its placement decisions, but that is outside the"},{"line_number":59,"context_line":"scope of this spec."},{"line_number":60,"context_line":""},{"line_number":61,"context_line":""},{"line_number":62,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":8,"id":"3afb71cf_2fdda642","line":59,"updated":"2015-06-05 17:04:50.000000000","message":"OK, probably wise.","commit_id":"288c1ce70b53010512a90c41a5811baf1fa518e0"},{"author":{"_account_id":1063,"name":"Ed Leafe","email":"ed@leafe.com","username":"ed-leafe"},"change_message_id":"9681eb5e8343cce4da25cbdb88cee0a9bbe8d9e8","unresolved":false,"context_lines":[{"line_number":56,"context_line":""},{"line_number":57,"context_line":"Information about cells will need to be fed into the scheduler in order for it"},{"line_number":58,"context_line":"to account for that during its placement decisions, but that is outside the"},{"line_number":59,"context_line":"scope of this spec."},{"line_number":60,"context_line":""},{"line_number":61,"context_line":""},{"line_number":62,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":8,"id":"1af86dd1_52609335","line":59,"in_reply_to":"3afb71cf_2fdda642","updated":"2015-06-12 13:59:39.000000000","message":"Agree that it is out of scope for this spec, but it is still a big question that needs to be answered.","commit_id":"288c1ce70b53010512a90c41a5811baf1fa518e0"}],"specs/liberty/approved/cells-scheduling.rst":[{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"6ca0b2520288df74d75e9328037ae83fabccfda7","unresolved":false,"context_lines":[{"line_number":20,"context_line":"When Nova is partitioned into cells, in order to boot an instance a scheduling"},{"line_number":21,"context_line":"decision will need to be made on both a cell and host to use.  And the decision"},{"line_number":22,"context_line":"of which cell to use will need to be made before the build process can be sent"},{"line_number":23,"context_line":"to a cell."},{"line_number":24,"context_line":""},{"line_number":25,"context_line":"Use Cases"},{"line_number":26,"context_line":"----------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_9e5ec010","line":23,"updated":"2015-04-15 10:37:41.000000000","message":"I have a feeling we need a separate spec to discuss how we make this scale inside the current nova-scheduler.\n\nMy current thinking is something like this:\n* small deployments have a single scheduler for all cells\n\nbut for larger deployments, roughly \u003e10k hosts:\n* assume there is a clean partition of requests and cells, i.e. a flavor tells you which subset of cells to look at\n* run a separate scheduler for each partition of cells\n* in the scheduler client, use a cells filter run to pick which which partition to build in, if two partitions are picked, fail the build\n* once the partition is picked, make the RPC call to that cell.\n\nAll the schedulers could probably benefit from the use of some ideas from the no-db scheduler, something like this:\n* the main issues is reading the full list of host state from the DB\n* caching that host state leads to excessive races between multiple scheduler process\n* instead of getting a list of all hosts before making each scheduling call, just get all the updates since you last checked\n* to implement this, we can use Kafaka to store updates from compute nodes\n* this assumes the scheduler is able to hold the full state of the host resources in memory\n* if the scheduler is restarted, we just go back to the beginning of the kafaka queue, and read all the messages\n* kafaka has a built in system that is able to delete old messages efficiently, effectively \"expiring\" them in order. We can use the window in a similar way to the current db service group timeout\n* the downside is that the host sends its full state to kafka every time, rather than partial updates, as it does today with the DB method\n* we could then go a step further... when we claim a resource from a compute node, we can wait until we get a message on the kafaka queue confirming the claim, so we know other schedulers are very unlikely to pick the same spot we just picked. and if the do, there is a quicker retry loop as the compute node would reject the resource claim if its expected state given by the scheduler doesn\u0027t match the currently known local state.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":20,"context_line":"When Nova is partitioned into cells, in order to boot an instance a scheduling"},{"line_number":21,"context_line":"decision will need to be made on both a cell and host to use.  And the decision"},{"line_number":22,"context_line":"of which cell to use will need to be made before the build process can be sent"},{"line_number":23,"context_line":"to a cell."},{"line_number":24,"context_line":""},{"line_number":25,"context_line":"Use Cases"},{"line_number":26,"context_line":"----------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_d5100951","line":23,"in_reply_to":"7aaa499b_08702b0a","updated":"2015-04-17 11:27:43.000000000","message":"On various IRC meetings, I think we have 2 concerns for cells :\n1/ provided we have a scalable scheduler, cells are just another abstraction layer for segregating a cloud, just like aggregates are \u003d\u003e that sounds like new cell-specific filters to me (and a new HostState field like host.cell)\n2/ the current scheduler is not scalable and it leaded to various efforts for workarounding that (like the CachingScheduler and the no-db-scheduler attempt).\n\nPoint #1 should be covered in that spec, but point #2 is scheduler-specific IMHO. Cells are just a new usecase for a scalable scheduler, but that shouldn\u0027t be discussed here.\n\nNot wanting to tease on #2 here, but I began to draft something on https://blueprints.launchpad.net/nova/+spec/shared-state-scheduler","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a65a328cb367e6adf039f43042eb1c6d1b832681","unresolved":false,"context_lines":[{"line_number":20,"context_line":"When Nova is partitioned into cells, in order to boot an instance a scheduling"},{"line_number":21,"context_line":"decision will need to be made on both a cell and host to use.  And the decision"},{"line_number":22,"context_line":"of which cell to use will need to be made before the build process can be sent"},{"line_number":23,"context_line":"to a cell."},{"line_number":24,"context_line":""},{"line_number":25,"context_line":"Use Cases"},{"line_number":26,"context_line":"----------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_08702b0a","line":23,"in_reply_to":"7aaa499b_9e5ec010","updated":"2015-04-15 14:54:00.000000000","message":"I\u0027m not sure I see the difference between the two level cells scheduler we currently have and your proposal for larger deployments.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"6ca0b2520288df74d75e9328037ae83fabccfda7","unresolved":false,"context_lines":[{"line_number":39,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":40,"context_line":""},{"line_number":41,"context_line":"The scheduler, not the current cells scheduler, will be updated to return a"},{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_83de670f","line":42,"updated":"2015-04-15 10:37:41.000000000","message":"Can\u0027t this just be (host, node)?\n\nThe node can have a property of cell_name to help find its location.\n\nReason why, the (host, node) identifies the resource you are selecting.\n\nThe cell is just a property of the host, just like an aggregate. It just happens that the cell defines which rabbit and db we will access to send the build request.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":39,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":40,"context_line":""},{"line_number":41,"context_line":"The scheduler, not the current cells scheduler, will be updated to return a"},{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_152551aa","line":42,"in_reply_to":"7aaa499b_034ecc3a","updated":"2015-04-17 11:27:43.000000000","message":"Agreed with John, there is no need to change the Scheduler API which returns an host object. That\u0027s the Cells Manager at the top API level which should look into the cell_host mapping table to see in which cell the returned host is part of.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a65a328cb367e6adf039f43042eb1c6d1b832681","unresolved":false,"context_lines":[{"line_number":39,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":40,"context_line":""},{"line_number":41,"context_line":"The scheduler, not the current cells scheduler, will be updated to return a"},{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_034ecc3a","line":42,"in_reply_to":"7aaa499b_83de670f","updated":"2015-04-15 14:54:00.000000000","message":"I see a cell as a scope when talking about (host, node), not a property of the host.\n\nMy understanding was that the scheduler was moving towards picking a node, not a host which is why I left it out.  And what we really care about is the cell not the host.\n\nBut I may need to understand the scheduling efforts a bit more to see if what I have is a bad direction.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"6ca0b2520288df74d75e9328037ae83fabccfda7","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_63cb5b1f","line":45,"updated":"2015-04-15 10:37:41.000000000","message":"-1 to that distinction\n\ncells filters and weights should be similar to aggregate filters and weights, and we need to fix how we implement both of those in the existing scheduler.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"ec1936744f28aafa841038faa627abf0b7b78776","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_ed5f883f","line":45,"updated":"2015-04-21 11:55:11.000000000","message":"Andrew, re: your usecase, we just need to make sure that operators (and upstream nova.conf) runs the related CellFilter as the first filter for the Scheduler","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"3f3bd1fb92d434bf92594c7a4ba67ddb2178f941","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_c79f315e","line":45,"in_reply_to":"3ab451fb_35a5ed03","updated":"2015-04-17 17:47:12.000000000","message":"I understand that.  But let\u0027s say I have 5 cells of 500 hosts each, and each cell has capabilities defined such that only one flavor can build in each and no flavor can build in two different cells.  Do I want to iterate over 2500 hosts in that first filter, or 5 cells?  When we can put hosts into buckets which have filterable properties it seems to me that it would be a good optimization to filter on the buckets before looking at the hosts in the buckets.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a65a328cb367e6adf039f43042eb1c6d1b832681","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_c318e420","line":45,"in_reply_to":"7aaa499b_63cb5b1f","updated":"2015-04-15 14:54:00.000000000","message":"This seems wrong to me.  If I wanted to write a cells filter I would need to iterate over every host which seems wasteful.  Why not just iterate over collections of hosts(cells)?  The same should probably be done for aggregates.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_35a5ed03","line":45,"in_reply_to":"7aaa499b_c318e420","updated":"2015-04-17 11:27:43.000000000","message":"Andrew, the current FilterScheduler design is O(n) for all the hosts but it actually refines the list of hosts being checked for each filter returning a subset of that list.\n\nSo, provided a CellFilter would return only the subset of hosts mapped to a specific cell, the other filters would just iterate over that subset.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"a6f369a7dfb7c480aaf2223f82788ee63391211b","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_31a144a4","line":45,"in_reply_to":"dac4157f_18866e65","updated":"2015-04-21 21:40:57.000000000","message":"I mostly agree that this is an orthogonal concern at the moment.  But I also want to figure out what we\u0027re going to do in the short term before we have a hash ring or other partitioning scheme in the scheduler.  But I realize that this doesn\u0027t belong in this spec.  We could have the same conversation about aggregates, rather than cells, on a purely scheduler spec.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"91157ea9340c1afe5a99fed556e268e3284dd78c","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_18866e65","line":45,"in_reply_to":"dac4157f_de3d8cbc","updated":"2015-04-21 16:24:37.000000000","message":"I think we still probably need to come out with a strategy of dividing your big fat cloud into smaller pieces for perf reasons, but I just think it\u0027s orthogonal.\nFor example, we can provide N sets of multiple schedulers able to scale out for 5 cells and provide an hash ring based on the compute node name for deciding which set of scheduler the compute reports to. That way, we would mix both worlds : a scaling-out scheduling system with an upper layer for ensuring isolation","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"eb4f5193f6ce7c0422d99b7b2aea656bcfd27ef8","unresolved":false,"context_lines":[{"line_number":42,"context_line":"(cell, node) tuple.  The input to the scheduler will remain unchanged.  There"},{"line_number":43,"context_line":"are currently some filters and weighers for the cells scheduler that will need"},{"line_number":44,"context_line":"to be ported into the scheduler.  We should keep a distinction in the scheduler"},{"line_number":45,"context_line":"between cells filters/weighers and host filter/weighers and each should be"},{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_de3d8cbc","line":45,"in_reply_to":"dac4157f_ed5f883f","updated":"2015-04-21 12:49:19.000000000","message":"Right, but should that cell filter operate on 2500 hosts, or 5 cells?  It\u0027s not much of a benefit if it still runs over all 2500 hosts.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"ec1936744f28aafa841038faa627abf0b7b78776","unresolved":false,"context_lines":[{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_cd5a8450","line":49,"updated":"2015-04-21 11:55:11.000000000","message":"Fair point","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":2033,"name":"Belmiro Moreira","email":"moreira.belmiro.email.lists@gmail.com","username":"moreira-belmiro-email-lists"},"change_message_id":"93c52682432c959fb68a7613078c3d98537cc8e5","unresolved":false,"context_lines":[{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"5aa745d1_ec511310","line":49,"updated":"2015-04-15 20:46:56.000000000","message":"I have some concerns about this spec because it doesn’t mention the possibility to have a two-tier scheduler configuration (cell and then host selection) even when using the same scheduler service but running at top and child level like today.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"3f3bd1fb92d434bf92594c7a4ba67ddb2178f941","unresolved":false,"context_lines":[{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_6794dd36","line":49,"in_reply_to":"3ab451fb_b590dd96","updated":"2015-04-17 17:47:12.000000000","message":"The goal at this point is to not require a one tier or two tier scheduler.  From a cells point of view we shouldn\u0027t care what the scheduler looks like.  All this is saying is that we should call a scheduler in this way and get this sort of response.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":46,"context_line":"given easy access to information that is relevant to them.  So they likely will"},{"line_number":47,"context_line":"not operate on the same data, though the cells scheduler might use an"},{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_b590dd96","line":49,"in_reply_to":"5aa745d1_ec511310","updated":"2015-04-17 11:27:43.000000000","message":"The goal is to have a global scheduler and not a 2-tier scheduler (one for the parent cell and one for the child cell). It just complicates the deployment and makes things harder to debug IMHO.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"a2cc217b1b431c0b0bafcf912b2e1297bd11436f","unresolved":false,"context_lines":[{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""},{"line_number":53,"context_line":""},{"line_number":54,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_f0ef73ae","line":51,"updated":"2015-04-13 19:47:42.000000000","message":"While I agree with you on providing the cells filters to the vanilla scheduler, I disagree with the interface : the scheduler should still return a compute node as a destination, and then the API would just have to call the related cell.\n\nThat way, providing some cell affinity is just orthogonal to providing some aggregate affinity for example, and would help the end-users to just do the cloud like they want (two-tier parent/child cells or more like grand-parent/parent/child)\n\nthat said, it leaves the main problem of a scalable scheduler unresolved. As Jay said, we have a big problem by getting the scheduler able to scale out, so I\u0027m advocating to see the cells effort join the scheduler effort for Lemming to see how we can resolve it.\n\nJust to be clear, I would like to see the Scheduler becoming an optimistic with shared-state scheduler like Google Omega as it\u0027s provided in this paper http://eurosys2013.tudos.org/wp-content/uploads/2013/paper/Schwarzkopf.pdf but that\u0027s not related to the cells work IMHO.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":48,"context_line":"aggregate of host data."},{"line_number":49,"context_line":""},{"line_number":50,"context_line":"During the transition period of moving to cells it should be possible to"},{"line_number":51,"context_line":"disable the cells portion of the scheduler."},{"line_number":52,"context_line":""},{"line_number":53,"context_line":""},{"line_number":54,"context_line":"Alternatives"}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_35f3ad0c","line":51,"in_reply_to":"7aaa499b_f0ef73ae","updated":"2015-04-17 11:27:43.000000000","message":"Sounds like I\u0027m repeating myself...","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":2033,"name":"Belmiro Moreira","email":"moreira.belmiro.email.lists@gmail.com","username":"moreira-belmiro-email-lists"},"change_message_id":"93c52682432c959fb68a7613078c3d98537cc8e5","unresolved":false,"context_lines":[{"line_number":57,"context_line":"We could continue using a two tiered scheduler like cellsv1.  This creates more"},{"line_number":58,"context_line":"deployment complexity and creates an unnecessary coupling between the"},{"line_number":59,"context_line":"architecture of Nova and the scheduler."},{"line_number":60,"context_line":""},{"line_number":61,"context_line":"Data model impact"},{"line_number":62,"context_line":"-----------------"},{"line_number":63,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"5aa745d1_cc5777c9","line":60,"updated":"2015-04-15 20:46:56.000000000","message":"For smaller deployments I agree. But in my view the main advantage of cells for large deployments is to have the services distributed. This avoids the complexity of having a big centralized service.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":2033,"name":"Belmiro Moreira","email":"moreira.belmiro.email.lists@gmail.com","username":"moreira-belmiro-email-lists"},"change_message_id":"941e9fb37cb338f49fd9619c0a652b3655db7ad5","unresolved":false,"context_lines":[{"line_number":57,"context_line":"We could continue using a two tiered scheduler like cellsv1.  This creates more"},{"line_number":58,"context_line":"deployment complexity and creates an unnecessary coupling between the"},{"line_number":59,"context_line":"architecture of Nova and the scheduler."},{"line_number":60,"context_line":""},{"line_number":61,"context_line":"Data model impact"},{"line_number":62,"context_line":"-----------------"},{"line_number":63,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_9276ada2","line":60,"in_reply_to":"3ab451fb_755ed5ea","updated":"2015-04-22 16:03:12.000000000","message":"As Andrew said probably this discussion doesn\u0027t belong into this spec however to answer @Sylvain:\nIn my previous comments I wasn\u0027t trying to mention the scheduler race conditions. \nFor me a cell is the scheduler cluster not the entire cloud. We have cells in different data centers and we have different scheduler requirements in some of them. Is a scheduler with a global overview the best option for these usecases?","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":57,"context_line":"We could continue using a two tiered scheduler like cellsv1.  This creates more"},{"line_number":58,"context_line":"deployment complexity and creates an unnecessary coupling between the"},{"line_number":59,"context_line":"architecture of Nova and the scheduler."},{"line_number":60,"context_line":""},{"line_number":61,"context_line":"Data model impact"},{"line_number":62,"context_line":"-----------------"},{"line_number":63,"context_line":""}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_755ed5ea","line":60,"in_reply_to":"5aa745d1_cc5777c9","updated":"2015-04-17 11:27:43.000000000","message":"You are misunderstanding: that\u0027s not because the current scheduler is monlithic because of race conditions that a global scheduler couldn\u0027t be distributed. The idea here is to have multiple schedulers able to be requested separately but would each have a global overview of the whole cluster.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":782,"name":"John Garbutt","email":"john@johngarbutt.com","username":"johngarbutt"},"change_message_id":"6ca0b2520288df74d75e9328037ae83fabccfda7","unresolved":false,"context_lines":[{"line_number":88,"context_line":""},{"line_number":89,"context_line":"Scheduling performance could be impacted since there are two sets of"},{"line_number":90,"context_line":"filters/weights.  Current cells users already deal with this impact, though it"},{"line_number":91,"context_line":"is spread between two services currently."},{"line_number":92,"context_line":""},{"line_number":93,"context_line":"Other deployer impact"},{"line_number":94,"context_line":"---------------------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"7aaa499b_6308fb91","line":91,"updated":"2015-04-15 10:37:41.000000000","message":"Honestly, the biggest impact is the scheduler reading the full list of all hosts, and looping over the larger list.\n\nCorrect ordering of the filters should make it minimal impact that there are extra filters and weights.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"ec1936744f28aafa841038faa627abf0b7b78776","unresolved":false,"context_lines":[{"line_number":88,"context_line":""},{"line_number":89,"context_line":"Scheduling performance could be impacted since there are two sets of"},{"line_number":90,"context_line":"filters/weights.  Current cells users already deal with this impact, though it"},{"line_number":91,"context_line":"is spread between two services currently."},{"line_number":92,"context_line":""},{"line_number":93,"context_line":"Other deployer impact"},{"line_number":94,"context_line":"---------------------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_ad59905c","line":91,"updated":"2015-04-21 11:55:11.000000000","message":"I think we can discuss on how to scale out the scheduler and if we want to split the list of hosts between multiple schedulers, but IMHO that shouldn\u0027t be in that spec. Let\u0027s consider that having 5K+ hosts is not a problem for the scheduler (yeah, I love to live difficulty)","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"3f3bd1fb92d434bf92594c7a4ba67ddb2178f941","unresolved":false,"context_lines":[{"line_number":88,"context_line":""},{"line_number":89,"context_line":"Scheduling performance could be impacted since there are two sets of"},{"line_number":90,"context_line":"filters/weights.  Current cells users already deal with this impact, though it"},{"line_number":91,"context_line":"is spread between two services currently."},{"line_number":92,"context_line":""},{"line_number":93,"context_line":"Other deployer impact"},{"line_number":94,"context_line":"---------------------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_8704897e","line":91,"in_reply_to":"3ab451fb_d56fe990","updated":"2015-04-17 17:47:12.000000000","message":"I think we\u0027re overlooking a huge optimization, as outlined above, by requiring filters that don\u0027t care about hosts to look at every host.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":7166,"name":"Sylvain Bauza","email":"sbauza@redhat.com","username":"sbauza"},"change_message_id":"19dbba7739aec6bae3d5106ca257713cd80d6697","unresolved":false,"context_lines":[{"line_number":88,"context_line":""},{"line_number":89,"context_line":"Scheduling performance could be impacted since there are two sets of"},{"line_number":90,"context_line":"filters/weights.  Current cells users already deal with this impact, though it"},{"line_number":91,"context_line":"is spread between two services currently."},{"line_number":92,"context_line":""},{"line_number":93,"context_line":"Other deployer impact"},{"line_number":94,"context_line":"---------------------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"3ab451fb_d56fe990","line":91,"in_reply_to":"7aaa499b_6308fb91","updated":"2015-04-17 11:27:43.000000000","message":"++","commit_id":"27861ad838ee8603096364f06829b77316dd766b"},{"author":{"_account_id":5441,"name":"Andrew Laski","email":"andrew@lascii.com","username":"alaski"},"change_message_id":"eb4f5193f6ce7c0422d99b7b2aea656bcfd27ef8","unresolved":false,"context_lines":[{"line_number":88,"context_line":""},{"line_number":89,"context_line":"Scheduling performance could be impacted since there are two sets of"},{"line_number":90,"context_line":"filters/weights.  Current cells users already deal with this impact, though it"},{"line_number":91,"context_line":"is spread between two services currently."},{"line_number":92,"context_line":""},{"line_number":93,"context_line":"Other deployer impact"},{"line_number":94,"context_line":"---------------------"}],"source_content_type":"text/x-rst","patch_set":4,"id":"dac4157f_3e7da86e","line":91,"in_reply_to":"dac4157f_ad59905c","updated":"2015-04-21 12:49:19.000000000","message":"Fair, it doesn\u0027t need to be in this spec.  The main point I want to make is that I think we should be writing cells/aggregate filters that don\u0027t operate on hosts, and I don\u0027t have another good place to express that opinion so I put it here :)\n\nI\u0027ll move that stuff out.","commit_id":"27861ad838ee8603096364f06829b77316dd766b"}]}
