Enables fully concurrent snapshot operations:
* Snapshot create- and delete operations can be started in any order
* Delete operations wait for snapshot finalization to finish, are batched as much as possible to improve efficiency and once enqueued in the cluster state prevent new snapshots from starting on data nodes until executed
* We could be even more concurrent here in a follow-up by interleaving deletes and snapshots on a per-shard level. I decided not to do this for now since it seemed not worth the added complexity yet. Due to batching+deduplicating of deletes the pain of having a delete stuck behind a long -running snapshot seemed manageable (dropped client connections + resulting retries don't cause issues due to deduplication of delete jobs, batching of deletes allows enqueuing more and more deletes even if a snapshot blocks for a long time that will all be executed in essentially constant time (due to bulk snapshot deletion, deleting multiple snapshots is mostly about as fast as deleting a single one))
* Snapshot creation is completely concurrent across shards, but per shard snapshots are linearized for each repository as are snapshot finalizations
See updated JavaDoc and added test cases for more details and illustration on the functionality.
Some notes:
The queuing of snapshot finalizations and deletes and the related locking/synchronization is a little awkward in this version but can be much simplified with some refactoring. The problem is that snapshot finalizations resolve their listeners on the `SNAPSHOT` pool while deletes resolve the listener on the master update thread. With some refactoring both of these could be moved to the master update thread, effectively removing the need for any synchronization around the `SnapshotService` state. I didn't do this refactoring here because it's a fairly large change and not necessary for the functionality but plan to do so in a follow-up.
This change allows for completely removing any trickery around synchronizing deletes and snapshots from SLM and 100% does away with SLM errors from collisions between deletes and snapshots.
Snapshotting a single index in parallel to a long running full backup will execute without having to wait for the long running backup as required by the ILM/SLM use case of moving indices to "snapshot tier". Finalizations are linearized but ordered according to which snapshot saw all of its shards complete first
Many of the parameters we pass into this method were only used to
build the `SnapshotInfo` instance to write.
This change simplifies the signature. Also, it seems less error prone to build
`SnapshotInfo` in `SnapshotsService` isntead of relying on the fact that each repository
implementation will build the correct `SnapshotInfo`.
Removing these limits as they cause unnecessarily many object in the blob stores.
We do not have to worry about BwC of this change since we do not support any 3rd party
implementations of Azure or GCS.
Also, since there is no valid reason to set a different than the default maximum chunk size at this
point, removing the documentation (which was incorrect in the case of Azure to begin with) for the setting
from the docs.
Closes#56018
This PR introduces two new fields in to `RepositoryData` (index-N) to track the blob name of `IndexMetaData` blobs and their content via setting generations and uuids. This is used to deduplicate the `IndexMetaData` blobs (`meta-{uuid}.dat` in the indices folders under `/indices` so that new metadata for an index is only written to the repository during a snapshot if that same metadata can't be found in another snapshot.
This saves one write per index in the common case of unchanged metadata thus saving cost and making snapshot finalization drastically faster if many indices are being snapshotted at the same time.
The implementation is mostly analogous to that for shard generations in #46250 and piggy backs on the BwC mechanism introduced in that PR (which means this PR needs adjustments if it doesn't go into `7.6`).
Relates to #45736 as it improves the efficiency of snapshotting unchanged indices
Relates to #49800 as it has the potential of loading the index metadata for multiple snapshots of the same index concurrently much more efficient speeding up future concurrent snapshot delete
We don't need to switch to the generic or snapshot pool for loading
cached repository data (i.e. most of the time in normal operation).
This makes `executeConsistentStateUpdate` less heavy if it has to retry
and lowers the chance of having to retry in the first place.
Also, this change allowed simplifying a few other spots in the codebase
where we would fork off to another pool just to load repository data.
With the removal of mapping types and the immutability of FieldTypeLookup in #58162, we no longer
have any cause to compare MappedFieldType instances. This means that we can remove all equals
and hashCode implementations, and in addition we no longer need the clone implementations which
were required for equals/hashcode testing. This greatly simplifies implementing new MappedFieldTypes,
which will be particularly useful for the runtime fields project.
In order to ensure that we do not write a broken piece of `RepositoryData`
because the phyiscal repository generation was moved ahead more than one step
by erroneous concurrent writing to a repository we must check whether or not
the current assumed repository generation exists in the repository physically.
Without this check we run the risk of writing on top of stale cached repository data.
Relates #56911
This commit creates a new Gradle plugin to provide a separate task name
and source set for running YAML based REST tests. The only project
converted to use the new plugin in this PR is distribution/archives/integ-test-zip.
For which the testing has been moved to :rest-api-spec since it makes the most
sense and it avoids a small but awkward change to the distribution plugin.
The remaining cases in modules, plugins, and x-pack will be handled in followups.
This plugin is distinctly different from the plugin introduced in #55896 since
the YAML REST tests are intended to be black box tests over HTTP. As such they
should not (by default) have access to the classpath for that which they are testing.
The YAML based REST tests will be moved to separate source sets (yamlRestTest).
The which source is the target for the test resources is dependent on if this
new plugin is applied. If it is not applied, it will default to the test source
set.
Further, this introduces a breaking change for plugin developers that
use the YAML testing framework. They will now need to either use the new source set
and matching task, or configure the rest resources to use the old "test" source set that
matches the old integTest task. (The former should be preferred).
As part of this change (which is also breaking for plugin developers) the
rest resources plugin has been removed from the build plugin and now requires
either explicit application or application via the new YAML REST test plugin.
Plugin developers should be able to fix the breaking changes to the YAML tests
by adding apply plugin: 'elasticsearch.yaml-rest-test' and moving the YAML tests
under a yamlRestTest folder (instead of test)
Currently we are leaving the settings to default port range in the nio
and netty4 http server test. This has recently led to tests failing due
to what appears to be a port conflict with other processes. This commit
modifies these tests to use the test case helper method to generate port
ranges.
Fixes#58433 and #58296.
Restoring from a snapshot (which is a particular form of recovery) does not currently take recovery throttling into account
(i.e. the `indices.recovery.max_bytes_per_sec` setting). While restores are subject to their own throttling (repository
setting `max_restore_bytes_per_sec`), this repository setting does not allow for values to be configured differently on a
per-node basis. As restores are very similar in nature to peer recoveries (streaming bytes to the node), it makes sense to
configure throttling in a single place.
The `max_restore_bytes_per_sec` setting is also changed to default to unlimited now, whereas previously it was set to
`40mb`, which is the current default of `indices.recovery.max_bytes_per_sec`). This means that no behavioral change
will be observed by clusters where the recovery and restore settings were not adapted.
Relates https://github.com/elastic/elasticsearch/issues/57023
Co-authored-by: James Rodewig <james.rodewig@elastic.co>
* Replace compile configuration usage with api (#58451)
- Use java-library instead of plugin to allow api configuration usage
- Remove explicit references to runtime configurations in dependency declarations
- Make test runtime classpath input for testing convention
- required as java library will by default not have build jar file
- jar file is now explicit input of the task and gradle will ensure its properly build
* Fix compile usages in 7.x branch
Netty4HttpServerTransportTests has started to fail intermittently. It
seems like unexpected successful responses are being received when the
test is simulating errors. This commit adds logging to the test to
provide additional information when there is an unexpected success. It
also adds the logging to the nio http test.
Now that MappedFieldType no longer extends lucene's FieldType, we need to have a
way of getting the index information about a field necessary for building text queries,
building term vectors, highlighting, etc. This commit introduces a new TextSearchInfo
abstraction that holds this information, and a getTextSearchInfo() method to
MappedFieldType to make it available. Field types that do not support text search can
just return null here.
This allows us to remove the MapperService.getLuceneFieldType() shim method.
Fixes a bug in TextFieldMapper serialization when index is false, and adds a
base-class test to ensure that all field mappers are tested against all variations
with defaults both included and excluded.
Fixes#58188
This is currently used to set the indexVersionCreated parameter on FieldMapper.
However, this parameter is only actually used by two implementations, and clutters
the API considerably. We should just remove it, and use it directly in the
implementations that require it.
MappedFieldType is a combination of two concerns:
* an extension of lucene's FieldType, defining how a field should be indexed
* a set of query factory methods, defining how a field should be searched
We want to break these two concerns apart. This commit is a first step to doing this, breaking
the inheritance relationship between MappedFieldType and FieldType. MappedFieldType
instead has a series of boolean flags defining whether or not the field is searchable or
aggregatable, and FieldMapper has a separate FieldType passed to its constructor defining
how indexing should be done.
Relates to #56814
This commit adds an optional field, `description`, to all ingest processors
so that users can explain the purpose of the specific processor instance.
Closes#56000.
* Remove usage of deprecated testCompile configuration
* Replace testCompile usage by testImplementation
* Make testImplementation non transitive by default (as we did for testCompile)
* Update CONTRIBUTING about using testImplementation for test dependencies
* Fail on testCompile configuration usage
Almost every outbound message is serialized to buffers of 16k pagesize.
We were serializing these messages off the IO loop (and retaining the concrete message
instance as well) and would then enqueue it on the IO loop to be dealt with as soon as the
channel is ready.
1. This would cause buffers to be held onto for longer than necessary, causing less reuse on average.
2. If a channel was slow for some reason, not only would concrete message instances queue up for it, but also 16k of buffers would be reserved for each message until it would be written+flushed physically.
With this change, the serialization happens on the event loop which effectively limits the number of buffers that `N` IO-threads will ever use so long as messages are small and channels writable.
Also, this change dereferences the reference to the concrete outbound message as soon as it has been serialized to save some more on GC.
This reduces the GC time for a default PMC run by about 50% in experiments (3 nodes, 2G heap each, loopback ... obvious caveat is that GC isn't that heavy in the first place with recent changes but still a measurable gain).
I also expect it to be helpful for master node stability by causing less of a spike if master is e.g. hit by a large number of requests that are processed batched (e.g. shard snapshot status updates) and responded to in a short time frame all at once.
Obviously, the downside to this change is that it introduces more latency on the IO loop for the serialization. But since we read all of these messages on the IO loop as well I don't see it as much of a qualitative change really and the more predictable buffer use seems much more valuable relatively.
* Fix GCS Mock Behavior for Missing Bucket
We were throwing a 500 instead of a 404 for a missing bucket.
This would make yaml tests needlessly wait for multiple seconds, retrying
the 500 response with backoff, in the test checking behavior for missing buckets.
We don't need to hold on to the request body past the beginning of sending
the response. There is no need to keep a reference to it until after the response
has been sent fully and we can eagerly release it here.
Note, this can be optimized further to release the contents even earlier but for now
this is an easy increment to saving some memory on the IO pool.
A few relatively obvious issues here:
* We cannot run the different IT runs (large blob setting one and normal integ run) concurrently
* We need to set the dependency tasks up correctly for the large blob run so that it works in isolation
* We can't use the `localAddress` for the location header of the resumable upload
(this breaks in YAML tests because GCS is using a loopback port forward for the initial request and the
local address will be chosen as the actual Docker container host)
Closes#57026
Merging logic is currently split between FieldMapper, with its merge() method, and
MappedFieldType, which checks for merging compatibility. The compatibility checks
are called from a third class, MappingMergeValidator. This makes it difficult to reason
about what is or is not compatible in updates, and even what is in fact updateable - we
have a number of tests that check compatibility on changes in mapping configuration
that are not in fact possible.
This commit refactors the compatibility logic so that it all sits on FieldMapper, and
makes it called at merge time. It adds a new FieldMapperTestCase base class that
FieldMapper tests can extend, and moves the compatibility testing machinery from
FieldTypeTestCase to here.
Relates to #56814
Adds tracking for the API calls performed by the Azure Storage
underlying SDK. It relies on the ability to hook a request
listener into the OperationContext.
Backport of #56773
Elasticsearch requires that a HttpRequest abstraction be implemented
by http modules before server processing. This abstraction controls when
underlying resources are released. This commit moves this abstraction to
be created immediately after content aggregation. This change will
enable follow-up work including moving Cors logic into the server
package and tracking bytes as they are aggregated from the network
level.
Add tracking for regular and multipart uploads.
Regular uploads are categorized as PUT.
Multi part uploads are categorized as POST.
The number of documents created for the test #testRequestStats
have been increased so all upload methods are exercised.
Backport of #56826
Add tracking for multipart and resumable uploads for GoogleCloudStorage.
For resumable uploads only the last request is taken into account for
billing, so that's the only request that's tracked.
Backport of #56821
Fixes the fact that repository metadata with the same settings still results in
multiple settings instances being cached as well as leaking settings on closing
a repository.
Closes#56702
In most cases we are seeing a `PooledHeapByteBuf` here now. No need to
redundantly create an new `ByteBuffer` and single element array for it
here when we can just directly unwrap its internal `byte[]`.
This change aims to fix our setup in CI so that we can run 7.x in
FIPS 140 mode. The major issue that we have in 7.x and did not
have in master is that we can't use the diagnostic trust manager
in FIPS mode in Java 8 with SunJSSE in FIPS approved mode as it
explicitly disallows the wrapping of X509TrustManager.
Previous attempts like #56427 and #52211 focused on disabling the
setting in all of our tests when creating a Settings object or
on setting fips_mode.enabled accordingly (which implicitly disables
the diagnostic trust manager). The attempts weren't future proof
though as nothing would forbid someone to add new tests without
setting the necessary setting and forcing this would be very
inconvenient for any other case ( see
#56427 (comment) for the full argumentation).
This change introduces a runtime check in SSLService that overrides
the configuration value of xpack.security.ssl.diagnose.trust and
disables the diagnostic trust manager when we are running in Java 8
and the SunJSSE provider is set in FIPS mode.
Mapper.Builder currently has some complex generics on it to allow fluent builder
construction. However, the second parameter, a return type from the build() method,
is unnecessary, as we can use covariant return types. This commit removes this second
generic parameter.
This is another part of the breakup of the massive BuildPlugin. This PR
moves the code for configuring publications to a separate plugin. Most
of the time these publications are jar files, but this also supports the
zip publication we have for integ tests.
We never do any file IO or other blocking work on the transport threads
so no tangible benefit can be derived from using more threads than CPUs
for IO.
There are however significant downsides to using more threads than necessary
with Netty in particular. Since we use the default setting for
`io.netty.allocator.useCacheForAllThreads` which is `true` we end up
using up to `16MB` of thread local buffer cache for each transport thread.
Meaning we potentially waste CPUs * 16MB of heap for unnecessary IO threads in addition to obvious inefficiencies of artificially adding extra context switches.
Ignacio Vera
Armin Braun
Alan Woodward
Rene Groeschke
Jake Landis
Tim Brooks
Yannick Welsch
Tal Levy
Jun Ohtani
Tanguy Leroux
Mark Vieira
markharwood
Mark Tozzi
Francisco Fernández Castaño
Ioannis Kakavas
Ryan Ernst