This commit enables concurrent search execution in the DFS phase, which is going to improve resource usage as well as performance of knn queries which benefit from both concurrent rewrite and collection.
We will enable concurrent execution for the query phase in a subsequent commit. While this commit does not introduce parallelism for the query phase, it introduces offloading sequential computation to the newly introduced executor. This is true both for situations where a single slice needs to be searched, as well as scenarios where a specific request does not support concurrency (currently only DFS phase does regardless of the request). Sequential collection is not offloaded only if the request includes aggregations that don't support offloading: composite, nested and cardinality as their post collection method must be executed in the same thread as the collection or we'll trip a lucene assertion that verifies that doc_values are pulled and consumed from the same thread.
## Technical details
This commit introduces a secondary executor, used exclusively to execute the concurrent bits of search. The search threads are still the ones that coordinate the search (where the caller search will originate from), but the actual work will be offloaded to the newly introduced executor.
We are offloading not only parallel execution but also sequential execution, to make the workload more predictable, as it would be surprising to have bits of search executed in either of the two thread pools. Also, that would introduce the possibility to suddenly run a higher amount of heavy operations overall (some in the caller thread and some in the separate threads), which could overload the system as well as make sizing of thread pools more difficult.
Note that fetch, together with other actions, is still executed in the search thread pool. This commit does not make the search thread pool merely a coordinating only thread pool, It does so only for what concerns the IndexSearcher#search operation itself, which is though a big portion of the different phases of search API execution.
Given that the searcher blocks waiting for all tasks to be completed, we take a simple approach of introducing a thread pool executor that has the same size as the existing search thread pool but relies on an unbounded queue. This simplifies handling of thread pool queue and rejections. In fact, we'd like to guarantee that the secondary thread pool won't reject, and delegate queuing entirely to the search thread pool which is the entry point for every search operation anyway. The principle behind this is that if you got a slot in the search thread pool, you should be able to complete your search, and rather quickly.
As part of this commit we are also introducing the ability to cancel tasks that have not started yet, so that if any task throws an exception, other tasks are prevented from starting needless computation.
Relates to #80693
Relates to #90700
Today by default the `SEARCH_COORDINATION` pool is sized at half the
allocated processors, or five if there are more than ten CPUs. Yet, if
we scale up a node to have more than ten CPUs, we probably want to scale
up the number of search coordination threads to match. This commit
removes the limit of five threads.
In #92309 we have aligned the size of the `search` and the `get` thread
pool but the docs still contain the prior `get` thread pool size. With
this commit we also align the docs.
Relates #92309
* Update threadpool.asciidoc
Starting from 8.0 the value of the `node.processors` setting is bounded by the number of available
processors https://github.com/elastic/elasticsearch/pull/44894
* Update docs/reference/modules/threadpool.asciidoc
Co-authored-by: Adam Locke <adam.locke@elastic.co>
This commit adds support for floating point node.processors setting.
This is useful when the nodes run in an environment where the CPU
time assigned to the ES node process is limited (i.e. using cgroups).
With this change, the system would be able to size the thread pools
accordingly, in this case it would round up the provided setting
to the closest integer.
Changes can-match from a shard-level to a node-level action, which helps avoid an explosion of shard-level can-match
subrequests in clusters with many shards, that can cause stability issues. Also introduces a new search_coordination
thread pool to handle the sending and handling of node-level can-match requests.
* Add new thread pool for critical operations
* Split critical thread pool into read and write
* Add POJO to hold thread pool names
* Add tests for critical thread pools
* Add thread pools to data streams
* Update settings for security plugin
* Retrieve ExecutorSelector from SystemIndices where possible
* Use a singleton ExecutorSelector
Adds new snapshot meta pool that is used to speed up the get snapshots API
by making `SnapshotInfo` load in parallel. Also use this pool to load
`RepositoryData`.
A follow-up to this would expand the use of this pool to the snapshot status
API and make it run in parallel as well.
This commit adds a dedicated threadpool for system index write
operations. The dedicated resources for system index writes serves as
a means to ensure that user activity does not block important system
operations from occurring such as the management of users and roles.
This commit introduces a new thread pool, `system_read`, which is
intended for use by system indices for all read operations (get and
search). The `system_read` pool is a fixed thread pool with a maximum
number of threads equal to lesser of half of the available processors
or 5. Given the combination of both get and read operations in this
thread pool, the queue size has been set to 2000. The motivation for
this change is to allow system read operations to be serviced in spite
of the number of user searches.
In order to avoid a significant performance hit due to pattern matching
on all search requests, a new metadata flag is added to mark indices
as system or non-system. Previously created system indices will have
flag added to their metadata upon upgrade to a version with this
capability.
Additionally, this change also introduces a new class, `SystemIndices`,
which encapsulates logic around system indices. Currently, the class
provides a method to check if an index is a system index and a method
to find a matching index descriptor given the name of an index.
Relates #50251
Relates #37867
Some of these characters are special to Asciidoctor and they ruin the
rendering on this page. Instead, we use a macro to passthrough these
characters without Asciidoctor applying any subtitutions to them. This
commit then addresses some rendering issues in the thread pool docs.
Co-authored-by: James Rodewig <james.rodewig@elastic.co>
The use of available processors, the terminology, and the settings
around it have evolved over time. This commit cleans up some places in
the codes and in the docs to adjust to the current terminology.
Synced flush was a brilliant idea. It supports instant recoveries with a
quite small implementation. However, with the presence of sequence
numbers and retention leases, it is no longer needed. This change
removes it from 8.0.
Relates #5077
The `bulk` threadpool is now called `write`, but `bulk` is still
used in some examples. This commit fixes that.
Also, the only way `threadpool.bulk.write: 30` is a valid increase in the size
of this threadpool is if you have 29 processors, which is an odd number of
processors to have. This commit removes the "more threads" bit.
This commit adds a rest endpoint for freezing and unfreezing an index.
Among other cleanups mainly fixing an issue accessing package private APIs
from a plugin that got caught by integration tests this change also adds
documentation for frozen indices.
Note: frozen indices are marked as `beta` and available as a basic feature.
Relates to #34352
This commit renames the bulk thread pool to the write thread pool. This
is to better reflect the fact that the underlying thread pool is used to
execute any document write request (single-document index/delete/update
requests, and bulk requests).
With this change, we add support for fallback settings
thread_pool.bulk.* which will be supported until 7.0.0.
We also add a system property so that the display name of the thread
pool remains as "bulk" if needed to avoid breaking users.
Now that single-document indexing requests are executed on the bulk
thread pool the index thread pool is no longer needed. This commit
removes this thread pool from Elasticsearch.
We want to remove the index thread pool as it is no longer needed since
single-document indexing requests are executed as bulk requests
now. Analyze requests are also executed on the index thread pool though
and they need a thread pool to execute on. The bulk thread does not seem
like the right thread pool, let us keep that thread pool conceptually
for bulk requests and free for bulk requests. None of the existing
thread pools make sense for analyze requests either. The generic thread
pool would be a terrible choice since it has an unbounded queue and that
is a bad idea for user-facing APIs. This commit introduces a small by
default (size=1, queue_size=16) thread pool for analyze requests.
This PR adds a new thread pool type: `fixed_auto_queue_size`. This thread pool
behaves like a regular `fixed` threadpool, except that every
`auto_queue_frame_size` operations (default: 10,000) in the thread pool,
[Little's Law](https://en.wikipedia.org/wiki/Little's_law) is calculated and
used to adjust the pool's `queue_size` either up or down by 50. A minimum and
maximum is taken into account also. When the min and max are the same value, a
regular fixed executor is used instead.
The `SEARCH` threadpool is changed to use this new type of thread pool. However,
the min and max are both set to 1000, meaning auto adjustment is opt-in rather
than opt-out.
Resolves#3890
Thread pool settings are no longer dynamically updatable since
da74323141. This commit removes a leftover
note from the thread pool module docs that incorrectly states that
thread pool settings are dynamically updatable.
This commit refactors the handling of thread pool settings so that the
individual settings can be registered rather than registering the top
level group. With this refactoring, individual plugins must now register
their own settings for custom thread pools that they need, but a
dedicated API is provided for this in the thread pool module. This
commit also renames the prefix on the thread pool settings from
"threadpool" to "thread_pool". This enables a hard break on the settings
so that:
- some of the settings can be given more sensible names (e.g., the max
number of threads in a scaling thread pool is now named "max" instead
of "size")
- change the soft limit on the number of threads in the bulk and
indexing thread pools to a hard limit
- the settings names for custom plugins for thread pools can be
prefixed (e.g., "xpack.watcher.thread_pool.size")
- remove dynamic thread pool settings
Relates #18674
This commit actually bounds the size of the generic thread pool. The
generic thread pool was of type cached, a thread pool with an unbounded
number of workers and an unbounded work queue. With this commit, the
generic thread pool is now of type scaling. As such, the cached thread
pool type has been removed. By default, the generic thread pool is
constructed with a core pool size of four, a max pool size of 128 and
idle workers can be reaped after a keep-alive time of thirty seconds
expires. The work queue for this thread pool remains unbounded.
In #17198, we removed suggest transport action, which
used the `suggest` threadpool to execute requests. Now
`suggest` threadpool is unused and suggest requests are
executed on the `search` threadpool.
This commit forbids the changing of thread pool types for any thread
pool. The motivation here is that these are expert settings with
little practical advantage.
Closes#14294, relates #2509, relates #2858, relates #5152
