Same as https://github.com/elastic/elasticsearch/pull/43288 for GCS.
We don't need to do the bucket exists check before using the repo, that just needlessly
increases the necessary permissions for using the GCS repository.
We have various ways of copying between two streams and handling thread-local
buffers throughout the codebase. This commit unifies a number of them and
removes buffer allocations in many spots.
For OSS plugins that being with repository-*, integTest
task is now a no-op and all of the tests are now executed via a test,
yamlRestTest, javaRestTest, or internalClusterTest.
related: #56841
related: #59444
Due to complicated access checks (reads and writes execute in their own access context) on some repositories (GCS, Azure, HDFS), using a hard coded buffer size of 4k for restores was needlessly inefficient.
By the same token, the use of stream copying with the default 8k buffer size for blob writes was inefficient as well.
We also had dedicated, undocumented buffer size settings for HDFS and FS repositories. For these two we would use a 100k buffer by default. We did not have such a setting for e.g. GCS though, which would only use an 8k read buffer which is needlessly small for reading from a raw `URLConnection`.
This commit adds an undocumented setting that sets the default buffer size to `128k` for all repositories. It removes wasteful allocation of such a large buffer for small writes and reads in case of HDFS and FS repositories (i.e. still using the smaller buffer to write metadata) but uses a large buffer for doing restores and uploading segment blobs.
This should speed up Azure and GCS restores and snapshots in a non-trivial way as well as save some memory when reading small blobs on FS and HFDS repositories.
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
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
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>
* 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.
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
Backporting #56585 to 7.x branch.
Adds tracking for the API calls performed by the GoogleCloudStorage
underlying SDK. It hooks an HttpResponseInterceptor to the SDK
transport layer and does http request filtering based on the URI
paths that we are interested to track. Unfortunately we cannot hook
a wrapper into the ServiceRPC interface since we're using different
levels of abstraction to implement retries during reads
(GoogleCloudStorageRetryingInputStream).
To read from GCS repositories we're currently using Google SDK's official BlobReadChannel,
which issues a new request every 2MB (default chunk size for BlobReadChannel) using range
requests, and fully downloads the chunk before exposing it to the returned InputStream. This
means that the SDK issues an awfully high number of requests to download large blobs.
Increasing the chunk size is not an option, as that will mean that an awfully high amount of
heap memory will be consumed by the download process.
The Google SDK does not provide the right abstractions for a streaming download. This PR
uses the lower-level primitives of the SDK to implement a streaming download, similar to what
S3's SDK does.
Also closes#55505
Adds ranged read support for GCS repositories in order to enable searchable snapshot support
for GCS.
As part of this PR, I've extracted some of the test infrastructure to make sure that
GoogleCloudStorageBlobContainerRetriesTests and S3BlobContainerRetriesTests are covering
similar test (as I saw those diverging in what they cover)
* Fix security manager bug writing large blobs to GCS
This commit addresses a security manager permissions issue writing large
blobs (on the resumable upload path) to GCS. The underlying issue here
is that we need to wrap the close and write calls on the channel. It is
not enough to do this:
SocketAccess.doPrivilegedVoidIOException(
() -> Streams.copy(
inputStream,
Channels.newOutputStream(client().writer(blobInfo, writeOptions))));
This reason that this is not enough is because Streams#copy will be in
the stacktrace and it is not granted the security manager permissions
needed to close or write this channel. We only grant those permissions
to classes loaded in the plugin classloader, and Streams#copy is from
the parent classloader. This is why we must wrap the close and write
calls as privileged, to truncate the Streams#copy call out of the
stacktrace.
The reason that this issue is not caught in testing is because the size
of data that we use in testing is too small to trigger the large blob
resumable upload path. Therefore, we address this by adding a system
property to control the threshold, which we can then set in tests to
exercise this code path. Prior to rewriting the writeBlobResumable
method to wrap the close and write calls as privileged, with this
additional test, we are able to reproduce the security manager
permissions issue. After adding the wrapping, this test now passes.
* Fix forbidden APIs issue
* Remove leftover debugging
This is a simple naming change PR, to fix the fact that "metadata" is a
single English word, and for too long we have not followed general
naming conventions for it. We are also not consistent about it, for
example, METADATA instead of META_DATA if we were trying to be
consistent with MetaData (although METADATA is correct when considered
in the context of "metadata"). This was a simple find and replace across
the code base, only taking a few minutes to fix this naming issue
forever.
Upgrading the GCS SDK to the most recent version.
Adjusting (i.e. improving) the REST mock accordingly.
This should significantly boost performance by pulling in
https://github.com/googleapis/java-core/issues/86 in some cases.
Tests in GoogleCloudStorageBlobStoreRepositoryTests are known
to be flaky on JDK 8 (#51446, #52430 ) and we suspect a JDK
bug (https://bugs.openjdk.java.net/browse/JDK-8180754) that triggers
some assertion on the server side logic that emulates the Google
Cloud Storage service.
Sadly we were not able to reproduce the failures, even when using
the same OS (Debian 9, Ubuntu 16.04) and JDK (Oracle Corporation
1.8.0_241 [Java HotSpot(TM) 64-Bit Server VM 25.241-b07]) of
almost all the test failures on CI. While we spent some time fixing
code (#51933, #52431) to circumvent the JDK bug they are still flaky
on JDK-8. This commit mute these tests for JDK-8 only.
Close ##52906
The countdown didn't work well here because it only returns `true` once the countdown reaches `0`
but can on subsequent executions return `false` again if a countdown at `0` is counted down again,
leading to more than the expected number of simulated failures.
Closes#52607
* Add Blob Download Retries to GCS Repository
Exactly as #46589 (and kept as close to it as possible code wise so we can dry things up in a follow-up potentially) but for GCS.
Closes#52319
For small uploads (that can still be up to 5MB!) we needlessly
reading the `InputStream` into a BAOS which entailed allocating
the `byte[]` for the stream contents twice (because to `toByteArray` on the BAOS copies).
Also, for resumeable uploads we were needlessly wrapping the output channel and running each individual write in its own privileged context when we could just wrap the whole upload in a single privileged context.
Relates #51593
This test was still very GC heavy in Java 8 runs in particular
which seems to slow down request processing to the point of timeouts
in some runs.
This PR completely removes the large number of O(MB) `byte[]` allocations
that were happening in the mock http handler which cuts the allocation rate
by about a factor of 5 in my local testing for the GC heavy `testSnapshotWithLargeSegmentFiles`
run.
Closes#51446Closes#50754
It's impossible to tell why #50754 fails without this change.
We're failing to close the `exchange` somewhere and there is no
write timeout in the GCS SDK (something to look into separately)
only a read timeout on the socket so if we're failing on an assertion without
reading the full request body (at least into the read-buffer) we're locking up
waiting forever on `write0`.
This change ensure the `exchange` is closed in the tests where we could lock up
on a write and logs the failure so we can find out what broke #50754.
* Remove BlobContainer Tests against Mocks
Removing all these weird mocks as asked for by #30424.
All these tests are now part of real repository ITs and otherwise left unchanged if they had
independent tests that didn't call the `createBlobStore` method previously.
The HDFS tests also get added coverage as a side-effect because they did not have an implementation
of the abstract repository ITs.
Closes#30424
* Remove Unused Single Delete in BlobStoreRepository
There are no more production uses of the non-bulk delete or the delete that throws
on missing so this commit removes both these methods.
Only the bulk delete logic remains. Where the bulk delete was derived from single deletes,
the single delete code was inlined into the bulk delete method.
Where single delete was used in tests it was replaced by bulk deleting.
Batch deletes get a response for every delete request, not just those that actually hit an existing blob.
The fact that we only responded for existing blobs leads to a degenerate response that throws a parse exception if a batch delete only contains non-existant blobs.
* Make BlobStoreRepository Aware of ClusterState (#49639)
This is a preliminary to #49060.
It does not introduce any substantial behavior change to how the blob store repository
operates. What it does is to add all the infrastructure changes around passing the cluster service to the blob store, associated test changes and a best effort approach to tracking the latest repository generation on all nodes from cluster state updates. This brings a slight improvement to the consistency
by which non-master nodes (or master directly after a failover) will be able to determine the latest repository generation. It does not however do any tricky checks for the situation after a repository operation
(create, delete or cleanup) that could theoretically be used to get even greater accuracy to keep this change simple.
This change does not in any way alter the behavior of the blobstore repository other than adding a better "guess" for the value of the latest repo generation and is mainly intended to isolate the actual logical change to how the
repository operates in #49060
All the implementations of `EsBlobStoreTestCase` use the exact same
bootstrap code that is also used by their implementation of
`EsBlobStoreContainerTestCase`.
This means all tests might as well live under `EsBlobStoreContainerTestCase`
saving a lot of code duplication. Also, there was no HDFS implementation for
`EsBlobStoreTestCase` which is now automatically resolved by moving the tests over
since there is a HDFS implementation for the container tests.
Same as #49518 pretty much but for GCS.
Fixing a few more spots where input stream can get closed
without being fully drained and adding assertions to make sure
it's always drained.
Moved the no-close stream wrapper to production code utilities since
there's a number of spots in production code where it's also useful
(will reuse it there in a follow-up).
This commit fixes the server side logic of "List Objects" operations
of Azure and S3 fixtures. Until today, the fixtures were returning a "
flat" view of stored objects and were not correctly handling the
delimiter parameter. This causes some objects listing to be wrongly
interpreted by the snapshot deletion logic in Elasticsearch which
relies on the ability to list child containers of BlobContainer (#42653)
to correctly delete stale indices.
As a consequence, the blobs were not correctly deleted from the
emulated storage service and stayed in heap until they got garbage
collected, causing CI failures like #48978.
This commit fixes the server side logic of Azure and S3 fixture when
listing objects so that it now return correct common blob prefixes as
expected by the snapshot deletion process. It also adds an after-test
check to ensure that tests leave the repository empty (besides the
root index files).
Closes#48978
Similarly to what has be done for Azure in #48636, this commit
adds a new :test:fixtures:gcs-fixture project which provides two
docker-compose based fixtures that emulate a Google Cloud
Storage service.
Some code has been extracted from existing tests and placed
into this new project so that it can be easily reused in other
projects.
This commit changes the test so that each node use a specific
service account and private key. It also changes how unique
request ids are generated for refresh token request using the
token itself, so that error count will be specific per node (each
node should execute a single refresh token request as tokens
are valid for 1 hour).
Similarly to what has been done for S3 in #45383, this commit
adds unit tests that verify the behavior of the SDK client and
blob container implementation for Google Storage when the
remote service returns errors.
The main purpose was to add an extra test to the specific retry
logic for 410-Gone errors added in #45963.
Relates #45963
GoogleCloudStorageBlobStore.deleteBlobsIgnoringIfNotExists() does
not correctly catch StorageException thrown by batch.submit().
In the case a snapshot is deleted through BlobStoreRepository.deleteSnapshot()
a storage exception is not caught (only IOException are) so the deletion is
interrupted and indices cannot be cleaned up. The storage exception bubbles
up to SnapshotService.deleteSnapshotFromRepository() but the listener that
removes the deletion from the cluster state is not executed, leaving the
deletion in the cluster state.
This bug has been reported in #46772 where batch.submit() threw an
exception in the test testIndicesDeletedFromRepository and following
tests failed because a snapshot deletion was running.
Relates #46772
This commit adds support for resumable uploads to the internal HTTP
server used in GoogleCloudStorageBlobStoreRepositoryTests. This
way we can also test the behavior of the Google's client when the
service returns server errors in response to resumable upload requests.
The BlobStore implementation for GCS has the choice between 2
methods to upload a blob: resumable and multipart. In the current
implementation, the client executes a resumable upload if the blob
size is larger than LARGE_BLOB_THRESHOLD_BYTE_SIZE,
otherwise it executes a multipart upload. This commit makes this
logic overridable in tests, allowing to randomize the decision of
using one method or the other.
The commit add support for single request resumable uploads
and chunked resumable uploads (the blob is uploaded into multiple
2Mb chunks; each chunk being a resumable upload). For this last
case, this PR also adds a test testSnapshotWithLargeSegmentFiles
which makes it more probable that a chunked resumable upload is
executed.
A resumable upload session can fail on with a 410 error and should
be retried in that case. I added retrying twice using resetting of
the given `InputStream` as the retry mechanism since the same
approach is used by the AWS S3 SDK already as well and relied upon
by the S3 repository implementation.
Related GCS documentation:
https://cloud.google.com/storage/docs/json_api/v1/status-codes#410_Gone
This commit modifies the HTTP server used in
GoogleCloudStorageBlobStoreRepositoryTests so that it randomly
returns server errors. The test does not inject server errors for the
following types of request: batch request, resumable upload request.
This commit removes the usage of MockGoogleCloudStoragePlugin in
GoogleCloudStorageBlobStoreRepositoryTests and replaces it by a
HttpServer that emulates the Storage service. This allows the repository
tests to use the real Google's client under the hood in tests and will allow
us to test the behavior of the snapshot/restore feature for GCS repositories
by simulating random server-side internal errors.
The HTTP server used to emulate the Storage service is intentionally simple
and minimal to keep things understandable and maintainable. Testing full
client options on the server side (like authentication, chunked encoding
etc) remains the responsibility of the GoogleCloudStorageFixture.
This commit starts from the simple premise that the use of node settings
in blob store repositories is a mistake. Here we see that the node
settings are used to get default settings for store and restore throttle
rates. Yet, since there are not any node settings registered to this
effect, there can never be a default setting to fall back to there, and
so we always end up falling back to the default rate. Since this was the
only use of node settings in blob store repository, we move them. From
this, several places fall out where we were chaining settings through
only to get them to the blob store repository, so we clean these up as
well. That leaves us with the changeset in this commit.
* Repository Cleanup Endpoint (#43900)
* Snapshot cleanup functionality via transport/REST endpoint.
* Added all the infrastructure for this with the HLRC and node client
* Made use of it in tests and resolved relevant TODO
* Added new `Custom` CS element that tracks the cleanup logic.
Kept it similar to the delete and in progress classes and gave it
some (for now) redundant way of handling multiple cleanups but only allow one
* Use the exact same mechanism used by deletes to have the combination
of CS entry and increment in repository state ID provide some
concurrency safety (the initial approach of just an entry in the CS
was not enough, we must increment the repository state ID to be safe
against concurrent modifications, otherwise we run the risk of "cleaning up"
blobs that just got created without noticing)
* Isolated the logic to the transport action class as much as I could.
It's not ideal, but we don't need to keep any state and do the same
for other repository operations
(like getting the detailed snapshot shard status)
Armin Braun
Jake Landis
Yannick Welsch
markharwood
Tanguy Leroux
Francisco Fernández Castaño
Ignacio Vera
Jason Tedor
Lee Hinman
Mark Vieira
Andrey Ershov