Today we choose to initialize max_seq_no_of_updates on primaries only so
we can deal with a situation where a primary is on an old node (before
6.5) which does not have MUS while replicas on new nodes (6.5+).
However, this strategy is quite complex and can lead to bugs (for
example #40249) since we have to assign a correct value (not too low) to
MSU in all possible situations (before recovering from translog,
restoring history on promotion, and handing off relocation).
Fortunately, we don't have to deal with this BWC in 7.0+ since all nodes
in the cluster should have MSU. This change simplifies the
initialization of MSU by always assigning it a correct value in the
constructor of Engine regardless of whether it's a replica or primary.
Relates #33842
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
This commit also backports the following commit:
Handle WRAP ops during SSL read
It is possible that a WRAP operation can occur while decrypting
handshake data in TLS 1.3. The SSLDriver does not currently handle this
well as it does not have access to the outbound buffer during read call.
This commit moves the buffer into the Driver to fix this issue. Data
wrapped during a read call will be queued for writing after the read
call is complete.
A stuck peer recovery in #40913 reveals that we indefinitely retry on
new cluster states if indexing translog operations hits a mapper
exception. We should not wait and retry if the mapping on the target is
as recent as the mapping that the primary used to index the replaying
operations.
Relates #40913
Motivated by slow snapshot deletes reported in e.g. #39656 and the fact that these likely are a contributing factor to repositories accumulating stale files over time when deletes fail to finish in time and are interrupted before they can complete.
* Makes snapshot deletion async and parallelizes some steps of the delete process that can be safely run concurrently via the snapshot thread poll
* I did not take the biggest potential speedup step here and parallelize the shard file deletion because that's probably better handled by moving to bulk deletes where possible (and can still be parallelized via the snapshot pool where it isn't). Also, I wanted to keep the size of the PR manageable.
* See https://github.com/elastic/elasticsearch/pull/39656#issuecomment-470492106
* Also, as a side effect this gives the `SnapshotResiliencyTests` a little more coverage for master failover scenarios (since parallel access to a blob store repository during deletes is now possible since a delete isn't a single task anymore).
* By adding a `ThreadPool` reference to the repository this also lays the groundwork to parallelizing shard snapshot uploads to improve the situation reported in #39657
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
* Introduce Delegating ActionListener Wrappers
* Dry up use cases of ActionListener that simply pass through the response or exception to another listener
hamcrest has some improvements in newer versions, like FileMatchers
that make assertions regarding file exists cleaner. This commit upgrades
to the latest version of hamcrest so we can start using new and improved
matchers.
* fix#35262 define deprecations of API's as a whole and urls
* document hot threads deprecated paths
* deprecate scroll_id as part of the URL, documented only as part of the body which is a safer behaviour as well
* use version numbers up to patch version
* rest spec parser picks up deprecated paths as paths too
(cherry picked from commit 7e06023e7603b7584bfd9ee4e8a1ccd82c208ce7)
It can be the case that while we are setting up expectations that also a
log message is appended. For example, if we are setting up these
expectations after a cluster has formed and messages start being sent
around the cluster. In this case, we would hit a concurrent modification
exception while we are mutating the expectations, and also while the
expectations are being iterated over as a message is appended. This
commit avoids this by using a copy-on-write array list which is safe for
concurrent modification and iteration. Note that another possible
approach here is to use synchronized, but that seems unnecessary since
we don't appear to rely on messages that are sent while we are setting
up expectations. Rather, we are setting up some expectations and some
situation that we think will cause those expectations to be met. Using
copy-on-write array list here is nice since we avoid bottlenecking these
tests on synchronizing these methods.
`Node#close` is pretty hard to rely on today:
- it might swallow exceptions
- it waits for 10 seconds for threads to terminate but doesn't signal anything
if threads are still not terminated after 10 seconds
This commit makes `IOException`s propagated and splits `Node#close` into
`Node#close` and `Node#awaitClose` so that the decision what to do if a node
takes too long to close can be done on top of `Node#close`.
It also adds synchronization to lifecycle transitions to make them atomic. I
don't think it is a source of problems today, but it makes things easier to
reason about.
* Adds Bulk delete API to blob container
* Implement bulk delete API for S3
* Adjust S3Fixture to accept both path styles for bulk deletes since the S3 SDK uses both during our ITs
* Closes#40250
Today a new replica of a closed index does not have a safe commit
invariant when its engine is opened because we won't initialize the
global checkpoint on a recovering replica until the finalize step. With
this change, we can achieve that property by creating a new translog
with the global checkpoint from the primary at the end of phase 1.
A small refactoring that removes the primaryTerm field from ReplicasProxy and
instead passes it directly in to the methods that need it. Relates #40706.
This is a dependency of #39504
Motivation:
By refactoring `TransportShardBulkAction#shardOperationOnPrimary` to async, we enable using `DeterministicTaskQueue` based tests to run indexing operations. This was previously impossible since we were blocking on the `write` thread until the `update` thread finished the mapping update.
With this change, the mapping update will trigger a new task in the `write` queue instead.
This change significantly enhances the amount of coverage we get from `SnapshotResiliencyTests` (and other potential future tests) when it comes to tracking down concurrency issues with distributed state machines.
The logical change is effectively all in `TransportShardBulkAction`, the rest of the changes is then simply mechanically moving the caller code and tests to being async and passing the `ActionListener` down.
Since the move to async would've added more parameters to the `private static` steps in this logic, I decided to inline and dry up (between delete and update) the logic as much as I could instead of passing the listener + wait-consumer down through all of them.
* Replace usages RandomizedTestingTask with built-in Gradle Test (#40978)
This commit replaces the existing RandomizedTestingTask and supporting code with Gradle's built-in JUnit support via the Test task type. Additionally, the previous workaround to disable all tasks named "test" and create new unit testing tasks named "unitTest" has been removed such that the "test" task now runs unit tests as per the normal Gradle Java plugin conventions.
(cherry picked from commit 323f312bbc829a63056a79ebe45adced5099f6e6)
* Fix forking JVM runner
* Don't bump shadow plugin version
Many gradle projects specifically use the -try exclude flag, because
there are many cases where auto-closeable resource ignore is never
referenced in body of corresponding try statement. Suppressing this
warning specifically in each case that it happens using
`@SuppressWarnings("try")` would be very verbose.
This change removes `-try` from any gradle project and adds it to the
build plugin. Also this change removes exclude flags from gradle projects
that is already specified in build plugin (for example -deprecation).
Relates to #40366
Today we are running our internal tests with bootstrap.memory_lock
enabled. This is not out default setting, and not the recommended
value. This commit switches to use the default value, which is to not
enable bootstrap.memory_lock.
We discussed recently that the cluster state API should be considered
"internal" and therefore our usual cast-iron stability guarantees do not hold
for this API.
However, there are a good number of REST tests that try to identify the master
node. Today they call `GET /_cluster/state` API and extract the master node ID
from the response. In fact many of these tests just want an arbitary node ID
(or perhaps a data node ID) so an alternative is to call `GET _nodes` or `GET
_nodes/data:true` and obtain a node ID from the keys of the `nodes` map in the
response.
This change adds the ability for YAML-based REST tests to extract an arbitrary
key from a map so that they can obtain a node ID from the nodes info API
instead of using the master node ID from the cluster state API.
Relates #40047.
This commit adds an InboundHandler to handle inbound message processing.
With this commit, this code is moved out of the TcpTransport.
Additionally, finer grained unit tests are added to ensure that the
inbound processing works as expected
Replicated closed indices can't be indexed into or searched, and therefore don't need a shard with
full indexing and search capabilities allocated. We can save on a lot of heap memory for those
indices by not allocating a mapper service and caching infrastructure (which preallocates a constant
amount per instance). Before this change, a 1GB ES instance could host 250 replicated closed
metricbeat indices (each index with one shard). After this change, the same instance can host 7300
replicated closed metricbeat instances (not that this would be a recommended configuration). Most
of the remaining memory is in the cluster state and the IndexSettings object.
Switches "discovery.type: single-node" from using a separate implementation for single-node discovery to using the existing standard discovery implementation, with two small adaptions:
- auto-bootstrapping, but requiring initial_master_nodes not to be set.
- not actively pinging other nodes using the Peerfinder
- not allowing other nodes to join its single-node cluster (if they have e.g. been set up using regular discovery and connect to the single-disco node).
Currently there are some components of message serializer and sending
that still occur in TcpTransport. This commit makes it possible to
send a message without the TcpTransport by moving all of the remaining
application logic to the OutboundHandler. Additionally, it adds unit
tests to ensure that this logic works as expected.
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
Additionally this commit back ports #40237
Remove Tracer from MockTransportService
Currently the TransportMessageListener is applied and used in the
Transport class. However, local requests and responses never make it to
this class. This PR moves the listener add/remove methods to the
TransportService. After this change the Transport can only have one
listener set with it. This one listener is the TransportService, which
will then propogate the events to the external listeners.
FilterDirectory.getPendingDeletions does not delegate, fixed
temporarily by overriding in StoreDirectory.
This in turn caused duplicate file name use after a trimUnsafeCommits
had been done, since a new IndexWriter would not consider the pending
deletes in IndexFileDeleter. This should only happen on windows (AFAIK).
Reenabled doing index updates for all tests using
IndexShardTests.indexOnReplicaWithGaps (which could fail due to above
when using mocked WindowsFS).
Added getPendingDeletions delegation to all elasticsearch
FilterDirectory subclasses that were not trivial test-only overrides to
minimize the risk of hitting this issue in another case.
* Mistake was made in #39662
* The response deserialized here is `org.elasticsearch.action.admin.cluster.snapshots.get.GetSnapshotsResponse` which uses `org.elasticsearch.snapshots.SnapshotInfo` which uses `org.elasticsearch.snapshots.SnapshotState` and not the shard state
If a replica were first reset due to one primary failover and then
promoted (before resync completes), its MSU would not include changes
since global checkpoint, leading to errors during translog replay.
Fixed by re-initializing MSU before restoring local history.
Unlike index operations which can fail at the document level to
analyzing errors, delete operations should never fail at the document
level whether soft-deletes is enabled or not. With this change, we will
always fail the engine if we fail to apply a delete operation to Lucene.
Closes#33256
We introduced WAIT_CLUSTERSTATE action in #19287 (5.0), but then stopped
using it since #25692 (6.0). This change removes that action and related
code in 7.x and 8.0.
Relates #19287
Relates #25692
When the method ensureGreen in QA tests is timed out, it does not
provide enough info for us to investigate why the testing index is
not green yet. With this change, we will dump the cluster state if
ensureGreen timed out.
Relates #32027
This PR introduces AsyncRecoveryTarget which executes remote calls of
peer recovery asynchronously. In this change, we also add a new
assertion to ensure that method sendBatch, which sends a batch of
history operations in phase2, is never called recursively on the same
thread. This new assertion will also be used in method sendFileChunks.
This commit unmutes NetworkDisruptionIT.
It makes changes necessary for Zen2 - avoids usage of
autoMinMasterNodes and selects cluster size, such that there is no
need to call AddVotingExclusion.
This test also introduces refactors a single method
prepareDistruptedCluster to be used by both test methods.
Unfortunately, NetworkDisruption is broken and the
testNetworkPartitionRemovalRestoresConnections "is fixed" by
introducing assertBusy - #38348.
Relates #36205
Relates #38348
(cherry picked from commit 97707c7f892636e5b75c3df546b067414acb27cd)
This change adds a wrapper for IndexSearcher that makes IndexSearcher#search(List, Weight, Collector) visible by
sub-classes. The wrapper is used by the ContextIndexSearcher to call this protected method on a searcher created by a plugin.
This ensures that an override of the protected method in an IndexSearcherWrapper plugin is called when a search is executed.
Closes#30758
Currently, we maintain a transport name ("mock-nio", "nio", "netty")
that is passed to a `TcpTransportChannel` when a request is received.
The value of this name is to associate with the task when we register a
task with the task manager. However, it is only possible to run ES with
one transport, so having an implementation specific name is unnecessary.
This commit removes the name and replaces it with the generic
"transport".
Today we load the shard history retention leases from disk whenever opening the
engine, and treat a missing file as an empty set of leases. However in some
cases this is inappropriate: we might be restoring from a snapshot (if the
target index already exists then there may be leases on disk) or
force-allocating a stale primary, and in neither case does it make sense to
restore the retention leases from disk.
With this change we write an empty retention leases file during recovery,
except for the following cases:
- During peer recovery the on-disk leases may be accurate and could be needed
if the recovery target is made into a primary.
- During recovery from an existing store, as long as we are not
force-allocating a stale primary.
Relates #37165
Computing the compressed size of the cluster state on every invocation
of cluster:monitor/state action is expensive, and the value of this
field is dubious anyway. Therefore we want to remove computing this
field. As a first step, we stop computing and return this field by
default. To avoid breaking users, we will give them a system property to
use to tide them over until the next major release when we will actually
remove this field. This comes with a deprecation warning too, and the
backport to the appropriate minor will also include a note in the
migration guide. There will be a follow-up to remove this field in the
next major version.
Today, when applying new cluster state we attempt to connect to all of its
nodes as a blocking part of the application process. This is the right thing to
do with new nodes, and is a no-op on any already-connected nodes, but is
questionable on known nodes from which we are currently disconnected: there is
a risk that we are partitioned from these nodes so that any attempt to connect
to them will hang until it times out. This can dramatically slow down the
application of new cluster states which hinders the recovery of the cluster
during certain kinds of partition.
If nodes are disconnected from the master then it is likely that they are to be
removed as part of a subsequent cluster state update, so there's no need to try
and reconnect to them like this. Moreover there is no need to attempt to
reconnect to disconnected nodes as part of the cluster state application
process, because we periodically try and reconnect to any disconnected nodes,
and handle their disconnectedness reasonably gracefully in the meantime.
This commit alters this behaviour to avoid reconnecting to known nodes during
cluster state application.
Resolves#29025.
Recently we have had a number of test issues related to blocking
activity occuring on the io thread. This commit adds a log warning for
when handling event takes a >150 milliseconds. This is implemented
for the MockNioTransport which is the transport used in
ESIntegTestCase.
* Back port build changes from #39102
This back-ports how versions are determined and bwc test are set up from
#39102 without enabling the bwc from current version tests so it's
easier/possible to backmerge future buld changes.
It's expected that the tets are lacking many of the required fixes in
this version to enable them.
* Wipe Snapshots Before Indices in RestTests
* If we have a snapshot ongoing from the previous test and enter this method, then deleting the indices fails, which in turn fails the whole wipe
* Fixed by first deleting/aborting snapshots
Today when a replicated write operation fails to execute on a replica,
the primary will reach out to the master to fail that replica (and mark
it stale). We then won't ack that request until the master removes the
failing replica; otherwise, we will lose the acked operation if the
failed replica is still in the in-sync set. However, if a node with the
primary is shutting down, we might ack such request even though we are
unable to send a shard-failure request to the master. This happens
because we ignore NodeClosedException which is triggered when the
ClusterService is being closed.
Closes#39467
Lucene added an optimization to leave the term dictionary on disk
for non-id like fields. This change happened very late in the release
processes such that it's better to have an escape hatch if certain
use-cases are hurt by this optimization. This setting might be
removed in the future if it turns out to be unnecessary.
When a node is joining the cluster we ensure that it can send requests to the
master _at that time_. If it joins the cluster and _then_ loses the ability to
send requests to the master then it should be removed from the cluster. Today
this is not the case: the master can still receive responses to its follower
checks, and receives acknowledgements to cluster state publications, so has no
reason to remove the node.
This commit changes the handling of follower checks so that they fail if they
come from a master that the other node was following but which it now believes
to have failed.
* Optimize Bulk Message Parsing and Message Length Parsing
* findNextMarker took almost 1ms per invocation during the PMC rally track
* Fixed to be about an order of magnitude faster by using Netty's bulk `ByteBuf` search
* It is unnecessary to instantiate an object (the input stream wrapper) and throw it away, just to read the `int` length from the message bytes
* Fixed by adding bulk `int` read to BytesReference
Nhat Nguyen
Tim Brooks
Armin Braun
Ryan Ernst
Martijn Laarman
Jason Tedor
Adrien Grand
David Turner
Mark Vieira
Martijn van Groningen
Yannick Welsch
Henning Andersen
Mayya Sharipova
Andrey Ershov
Jim Ferenczi
Alpar Torok
Simon Willnauer