)]}'
{"specs/newton/approved/probabilistic-weighted-placement-algorithm.rst":[{"author":{"_account_id":8688,"name":"Alexis Lee","email":"openstack@lxsli.co.uk","username":"lxsli"},"change_message_id":"f008acc7f2099a856317a5128b6e315f3b5da619","unresolved":false,"context_lines":[{"line_number":57,"context_line":"    pick_from_distribution(host_probabilties)"},{"line_number":58,"context_line":""},{"line_number":59,"context_line":"Instead of choosing randomly from the top ``n`` hosts the algorithm will assign"},{"line_number":60,"context_line":"a probability of being chosen to **every** host based on its weight. If"},{"line_number":61,"context_line":"a host has twice the weight of another it will be given twice the probability"},{"line_number":62,"context_line":"of being chosen. The host list will form a proper probability distribution:"},{"line_number":63,"context_line":"the sum of all probabilities across hosts will be 1."}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a122d0e_45aacc86","line":60,"updated":"2016-04-21 16:15:03.000000000","message":"The weights today only matter in terms of ordering. This would make the ratios interesting, which could require a lot more weight-tuning (which is already rather hated I believe).\n\nEG today I can ensure hostB enters the subset by multiplying its weight by either 100 or 1000000. Under this new algorithm, the latter would be a much stronger prioritisation.","commit_id":"ac95dca22a3481de918537e4777519dbd22c0290"},{"author":{"_account_id":18247,"name":"Mark Doffman","email":"mdoffman@linux.vnet.ibm.com","username":"doffm"},"change_message_id":"bd15983dc77b1cf78a47406a40c27917a6402f85","unresolved":false,"context_lines":[{"line_number":57,"context_line":"    pick_from_distribution(host_probabilties)"},{"line_number":58,"context_line":""},{"line_number":59,"context_line":"Instead of choosing randomly from the top ``n`` hosts the algorithm will assign"},{"line_number":60,"context_line":"a probability of being chosen to **every** host based on its weight. If"},{"line_number":61,"context_line":"a host has twice the weight of another it will be given twice the probability"},{"line_number":62,"context_line":"of being chosen. The host list will form a proper probability distribution:"},{"line_number":63,"context_line":"the sum of all probabilities across hosts will be 1."}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a122d0e_852f723d","line":60,"in_reply_to":"1a122d0e_45aacc86","updated":"2016-04-24 12:02:12.000000000","message":"Good catch. The system we have currently only takes the top N of an ordered list and as such will likely be much less sensitive to the weights ratios than this proposal. I will update the spec to discuss this. \n\nYes, the relative weights are a nightmare, there is no good way to decide on a value except LOTS of testing. Same goes for the host subset config variable. There is really NO advice on what this should be set to. \n\nI am currently working on a separate proposal to deal with the relative weights issue. Will add link when finished.","commit_id":"ac95dca22a3481de918537e4777519dbd22c0290"},{"author":{"_account_id":8688,"name":"Alexis Lee","email":"openstack@lxsli.co.uk","username":"lxsli"},"change_message_id":"f008acc7f2099a856317a5128b6e315f3b5da619","unresolved":false,"context_lines":[{"line_number":63,"context_line":"the sum of all probabilities across hosts will be 1."},{"line_number":64,"context_line":""},{"line_number":65,"context_line":"The new algorithm will be added under a config option that will, by default"},{"line_number":66,"context_line":"be disabled."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":"A proof of concept for this change is available on Gerrit. [1]_"},{"line_number":69,"context_line":""}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a122d0e_058e1427","line":66,"updated":"2016-04-21 16:15:03.000000000","message":"So the FilterScheduler will check this option?","commit_id":"ac95dca22a3481de918537e4777519dbd22c0290"},{"author":{"_account_id":18247,"name":"Mark Doffman","email":"mdoffman@linux.vnet.ibm.com","username":"doffm"},"change_message_id":"bd15983dc77b1cf78a47406a40c27917a6402f85","unresolved":false,"context_lines":[{"line_number":63,"context_line":"the sum of all probabilities across hosts will be 1."},{"line_number":64,"context_line":""},{"line_number":65,"context_line":"The new algorithm will be added under a config option that will, by default"},{"line_number":66,"context_line":"be disabled."},{"line_number":67,"context_line":""},{"line_number":68,"context_line":"A proof of concept for this change is available on Gerrit. [1]_"},{"line_number":69,"context_line":""}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a122d0e_a52c2e4b","line":66,"in_reply_to":"1a122d0e_058e1427","updated":"2016-04-24 12:02:12.000000000","message":"Yes.","commit_id":"ac95dca22a3481de918537e4777519dbd22c0290"},{"author":{"_account_id":8688,"name":"Alexis Lee","email":"openstack@lxsli.co.uk","username":"lxsli"},"change_message_id":"f008acc7f2099a856317a5128b6e315f3b5da619","unresolved":false,"context_lines":[{"line_number":183,"context_line":""},{"line_number":184,"context_line":"* Create the new placement algorithm. Write functions required to generate"},{"line_number":185,"context_line":"  a probability distribution from a weighted list of hosts."},{"line_number":186,"context_line":""},{"line_number":187,"context_line":"Dependencies"},{"line_number":188,"context_line":"\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d\u003d"},{"line_number":189,"context_line":""}],"source_content_type":"text/x-rst","patch_set":2,"id":"1a122d0e_05c61416","line":186,"updated":"2016-04-21 16:15:03.000000000","message":"For this to be useful, you also need to go through the existing weighers and ensure the weights used correctly reflect the factor\u0027s importance. Maybe make them tunable as operators are likely to disagree on this.","commit_id":"ac95dca22a3481de918537e4777519dbd22c0290"}]}