Move the terms index of _id off-heap.#52405
Merged
jpountz merged 4 commits intoelastic:masterfrom Feb 19, 2020
Merged
Conversation
In elastic#42838 we moved the terms index of all fields off-heap except the `_id` field because we were worried it might make indexing slower. In general, the indexing rate is only affected if explicit IDs are used, as otherwise Elasticsearch almost never performs lookups in the terms dictionary for the purpose of indexing. So it's quite wasteful to require the terms index of `_id` to be loaded on-heap for users who have append-only workloads. Furthermore I've been conducting benchmarks when indexing with explicit ids on the http_logs dataset that suggest that the slowdown is low enough that it's probably not worth forcing the terms index to be kept on-heap. Here are some numbers for the median indexing rate in docs/s: | Run | Master | Patch | | --- | ------- | ------- | | 1 | 45851.2 | 46401.4 | | 2 | 45192.6 | 44561.0 | | 3 | 45635.2 | 44137.0 | | 4 | 46435.0 | 44692.8 | | 5 | 45829.0 | 44949.0 | And now heap usage in MB for segments: | Run | Master | Patch | | --- | ------- | -------- | | 1 | 41.1720 | 0.352083 | | 2 | 45.1545 | 0.382534 | | 3 | 41.7746 | 0.381285 | | 4 | 45.3673 | 0.412737 | | 5 | 45.4616 | 0.375063 | Indexing rate decreased by 1.8% on average, while memory usage decreased by more than 100x. The `http_logs` dataset contains small documents and has a simple indexing chain. More complex indexing chains, e.g. with more fields, ingest pipelines, etc. would see an even lower decrease of indexing rate.
Collaborator
|
Pinging @elastic/es-distributed (:Distributed/Engine) |
Contributor
Author
|
I'm sharing all the Rally outputs for completeness: Master: Patch: |
dnhatn
approved these changes
Feb 18, 2020
| }, Property.IndexScope); | ||
| // This setting is undocumented as it is considered as an escape hatch. | ||
| public static final Setting<Boolean> ON_HEAP_ID_TERMS_INDEX = | ||
| Setting.boolSetting("index.force_memory_id_terms_dictinary", false, Property.IndexScope); |
Contributor
Author
There was a problem hiding this comment.
Thanks for catching this, you were not nit-picking!
jpountz
added a commit
to jpountz/elasticsearch
that referenced
this pull request
Feb 19, 2020
Backport of elastic#52405 In elastic#42838 we moved the terms index of all fields off-heap except the `_id` field because we were worried it might make indexing slower. In general, the indexing rate is only affected if explicit IDs are used, as otherwise Elasticsearch almost never performs lookups in the terms dictionary for the purpose of indexing. So it's quite wasteful to require the terms index of `_id` to be loaded on-heap for users who have append-only workloads. Furthermore I've been conducting benchmarks when indexing with explicit ids on the http_logs dataset that suggest that the slowdown is low enough that it's probably not worth forcing the terms index to be kept on-heap. Here are some numbers for the median indexing rate in docs/s: | Run | Master | Patch | | --- | ------- | ------- | | 1 | 45851.2 | 46401.4 | | 2 | 45192.6 | 44561.0 | | 3 | 45635.2 | 44137.0 | | 4 | 46435.0 | 44692.8 | | 5 | 45829.0 | 44949.0 | And now heap usage in MB for segments: | Run | Master | Patch | | --- | ------- | -------- | | 1 | 41.1720 | 0.352083 | | 2 | 45.1545 | 0.382534 | | 3 | 41.7746 | 0.381285 | | 4 | 45.3673 | 0.412737 | | 5 | 45.4616 | 0.375063 | Indexing rate decreased by 1.8% on average, while memory usage decreased by more than 100x. The `http_logs` dataset contains small documents and has a simple indexing chain. More complex indexing chains, e.g. with more fields, ingest pipelines, etc. would see an even lower decrease of indexing rate.
sbourke
pushed a commit
to sbourke/elasticsearch
that referenced
this pull request
Feb 19, 2020
In elastic#42838 we moved the terms index of all fields off-heap except the `_id` field because we were worried it might make indexing slower. In general, the indexing rate is only affected if explicit IDs are used, as otherwise Elasticsearch almost never performs lookups in the terms dictionary for the purpose of indexing. So it's quite wasteful to require the terms index of `_id` to be loaded on-heap for users who have append-only workloads. Furthermore I've been conducting benchmarks when indexing with explicit ids on the http_logs dataset that suggest that the slowdown is low enough that it's probably not worth forcing the terms index to be kept on-heap. Here are some numbers for the median indexing rate in docs/s: | Run | Master | Patch | | --- | ------- | ------- | | 1 | 45851.2 | 46401.4 | | 2 | 45192.6 | 44561.0 | | 3 | 45635.2 | 44137.0 | | 4 | 46435.0 | 44692.8 | | 5 | 45829.0 | 44949.0 | And now heap usage in MB for segments: | Run | Master | Patch | | --- | ------- | -------- | | 1 | 41.1720 | 0.352083 | | 2 | 45.1545 | 0.382534 | | 3 | 41.7746 | 0.381285 | | 4 | 45.3673 | 0.412737 | | 5 | 45.4616 | 0.375063 | Indexing rate decreased by 1.8% on average, while memory usage decreased by more than 100x. The `http_logs` dataset contains small documents and has a simple indexing chain. More complex indexing chains, e.g. with more fields, ingest pipelines, etc. would see an even lower decrease of indexing rate.
This was referenced Feb 20, 2020
bpintea
added a commit
that referenced
this pull request
Apr 20, 2020
mfussenegger
added a commit
to crate/crate
that referenced
this pull request
Apr 30, 2020
Port of elastic/elasticsearch#52405 but excludes the undocumented setting. Relates to #9796
mfussenegger
added a commit
to crate/crate
that referenced
this pull request
Apr 30, 2020
Port of elastic/elasticsearch#52405 but excludes the undocumented setting. Relates to #9796
5 tasks
mergify bot
pushed a commit
to crate/crate
that referenced
this pull request
Apr 30, 2020
Port of elastic/elasticsearch#52405 but excludes the undocumented setting. Relates to #9796
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
In #42838 we moved the terms index of all fields off-heap except the
_idfield because we were worried it might make indexing slower. Ingeneral, the indexing rate is only affected if explicit IDs are used, as
otherwise Elasticsearch almost never performs lookups in the terms
dictionary for the purpose of indexing. So it's quite wasteful to
require the terms index of
_idto be loaded on-heap for users who haveappend-only workloads. Furthermore I've been conducting benchmarks when
indexing with explicit ids on the http_logs dataset that suggest that
the slowdown is low enough that it's probably not worth forcing the terms
index to be kept on-heap. Here are some numbers for the median indexing
rate in docs/s:
And now heap usage in MB for segments:
Indexing rate decreased by 1.8% on average, while memory usage decreased
by more than 100x.
The
http_logsdataset contains small documents and has a simpleindexing chain. More complex indexing chains, e.g. with more fields,
ingest pipelines, etc. would see an even lower decrease of indexing rate.