summaryrefslogtreecommitdiffstats
path: root/container-search/src/test/java/com/yahoo/search/dispatch
Commit message (Collapse)AuthorAgeFilesLines
* more cleanup of MatchFeatureDataArne H Juul2021-11-091-8/+6
|
* less duplication of hashmapsArne H Juul2021-11-091-0/+104
| | | | | * instead of building an ObjectValue for each hit (each with its own mapping of names->values) use a shared Hashlet.
* Update 2020 Oath copyrights.gjoranv2021-10-272-2/+2
|
* Update 2019 Oath copyrights.gjoranv2021-10-273-3/+3
|
* remove unused RpcFill protocol implementationArne H Juul2021-10-194-263/+0
|
* Update Verizon Media copyright notices.gjoranv2021-10-073-3/+3
|
* Update 2019 Yahoo Holdings copyright notices.gjoranv2021-10-072-2/+2
|
* Update 2018 copyright notices.gjoranv2021-10-074-4/+4
|
* Update 2017 copyright notices.gjoranv2021-10-072-2/+2
|
* Merge branch 'master' into balder/do-not-depend-on-clusterinfoHenning Baldersheim2021-09-3013-62/+612
|\
| * Enable top-k optimization only if balanced and non-sparseJon Bratseth2021-07-021-24/+46
| |
| * Separate balanced and sparseJon Bratseth2021-07-021-2/+22
| |
| * Infer groupJon Bratseth2021-07-024-35/+21
| |
| * Allow deviation of at least 1 documentHarald Musum2021-06-301-0/+18
| | | | | | | | Let content be well-balanced when there are few docs in a cluster
| * Gc unused config options: min-group-coverage and max nodes per groupJon Bratseth2021-06-151-1/+0
| |
| * Revert "Revert "Don't consider number of working nodes in coverage""Jon Bratseth2021-05-111-2/+16
| |
| * Revert "Don't consider number of working nodes in coverage"Jon Bratseth2021-05-101-16/+2
| |
| * Don't consider number of working nodes in coverageJon Bratseth2021-05-101-2/+16
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Tring to figure out the right groups to send queries to based on the number of nodes in the group has many potential issues at times of topology changes. Since we could the number of documents available in each group by summing documents in working nodes, we do not need to also separately consider the number of working nodes in the group for correctness. Since we use adaptive dispatching by default we also do not need to consider it to avoid overloading groups with less resources available but enough documents.
| * Use median not average document count to determine group coverageJon Bratseth2021-04-158-8/+156
| | | | | | | | If a group has too many nodes, all others will have less than average.
| * Minor group logic improvementsJon Bratseth2021-03-311-1/+1
| | | | | | | | | | | | - Non-functional cleanup - Add cluster id to log messages - Don't warn that we only have N/N and require N nodes (all for the same N)
| * Add possibility to select a subset of groups and apply random selection ↵Henning Baldersheim2021-01-111-22/+39
| | | | | | | | among them.
| * Revert "Revert "Disable topk optimisation on dispatch when content ↵Henning Baldersheim2021-01-082-14/+62
| | | | | | | | distribution is se…""
| * Revert "Disable topk optimisation on dispatch when content distribution is ↵Henning Baldersheim2021-01-082-62/+14
| | | | | | | | se…"
| * Disable topk optimisation on dispatch when content distribution is severly ↵Henning Baldersheim2021-01-072-14/+62
| | | | | | | | | | | | | | | | skewed. When the skew is too large the assumption that docs are evenly and randomly distributed hold. The impact and is larger on smaller systems. In large systems the where this optimisation is more important, the probabilitity of large skew will be less.
| * Shutdown and join the rpc client transport threads.Henning Baldersheim2020-09-212-2/+5
| |
| * Use existing uncompressed lengthJon Bratseth2020-07-011-2/+2
| |
| * Formatting changes only#Jon Bratseth2020-07-011-4/+4
| |
| * Add support for serializing and compressing once, instead of once per ↵Henning Baldersheim2020-06-302-3/+13
| | | | | | | | backend node.
| * locale-independent testkkraune2020-06-181-3/+5
| |
| * Allow from 5% skew in document distribution and still get good results when ↵Henning Baldersheim2020-05-282-22/+35
| | | | | | | | asking for many hits.
| * Add the expected table for documentations and assert on changes.Henning Baldersheim2020-05-281-6/+34
| |
| * Add a test to show how topk affects as hits grows.Henning Baldersheim2020-05-271-0/+39
| |
| * Specify round-robin dispatch policy explisit as default has changed to adaptive.Henning Baldersheim2020-05-201-0/+1
| |
| * The estimate is not bounded by [1, K] so it must be capped.Henning Baldersheim2020-05-191-0/+58
| |
| * Propagate relevance even if sorting is used.Henning Baldersheim2020-04-161-4/+8
| |
| * isWorking is tri-stateJon Bratseth2020-04-161-1/+0
| |
| * Merge pull request #12918 from vespa-engine/balder/top-k-probabilityHenning Baldersheim2020-04-164-0/+58
| |\ | | | | | | Introduce top-k-probability and use it to fetch correct proper amount…
| | * Add copyright too.Henning Baldersheim2020-04-151-0/+1
| | |
| | * Add query control of top-k-probability.Henning Baldersheim2020-04-153-0/+38
| | |
| | * Make SearchCluster.TopKEstimator a top level class.Henning Baldersheim2020-04-152-12/+19
| | |
| | * Introduce top-k-probability and use it to fetch correct proper amount of ↵Henning Baldersheim2020-04-151-0/+12
| | | | | | | | | | | | hits from each partition
| * | Don't avoid group blocking feed when groups = 1Jon Bratseth2020-04-151-4/+16
| | |
| * | Avoid a single group rejecting feedJon Bratseth2020-04-153-10/+47
| |/
| * Revert "Revert "Revert "Revert "Don't take combined clusters of size 1 down""""Jon Bratseth2020-03-261-1/+15
| |
| * Revert "Revert "Revert "Don't take combined clusters of size 1 down"""Jon Bratseth2020-03-261-15/+1
| |
| * Revert "Revert "Don't take combined clusters of size 1 down""Jon Bratseth2020-03-251-1/+15
| |
| * Revert "Don't take combined clusters of size 1 down"Harald Musum2020-03-251-15/+1
| |
| * Don't take combined clusters of size 1 downJon Bratseth2020-03-251-1/+15
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This can lead to a deadlock: - host-admin needs to suspend node before it reduces the CPU allocation - suspension means setting storage node in maintenance, distributor down - cluster controller figures this means the cluster is down - the container on the same node (being a combined cluster) receives report from the downstream storage node of being offline, and changes its /state/v1/health to down - being a combined cluster node w/container, the host-admin must verify /health/v1/status is UP before allowing resume, which it isn't We have no good options when the content node is down and size is 1, and do not much care about availability in this case by definition, so keeping the container in rotation should be fine.
| * Revert "Revert "Revert "Revert "Create a resourcepool so that we do not need ↵Henning Baldersheim2020-02-193-6/+5
| | | | | | | | to reconnect to content …""""
| * Revert "Revert "Revert "Create a resourcepool so that we do not need to ↵Harald Musum2020-02-193-5/+6
| | | | | | | | reconnect to content …"""