Sequence BWC logic consists of two elements:
1) Wire level BWC using stream versions.
2) A changed to the global checkpoint maintenance semantics.
For the sequence number infra to work with a mixed version clusters, we have to consider situation where the primary is on an old node and replicas are on new ones (i.e., the replicas will receive operations without seq#) and also the reverse (i.e., the primary sends operations to a replica but the replica can't process the seq# and respond with local checkpoint). An new primary with an old replica is a rare because we do not allow a replica to recover from a new primary. However, it can occur if the old primary failed and a new replica was promoted or during primary relocation where the source primary is treated as a replica until the master starts the target.
1) Old Primary & New Replica - this case is easy as is taken care of by the wire level BWC. All incoming requests will have their seq# set to `UNASSIGNED_SEQ_NO`, which doesn't confuse the local checkpoint logic (keeping it at `NO_OPS_PERFORMED`)
2) New Primary & Old replica - this one is trickier as the global checkpoint service currently takes all in sync replicas into consideration for the global checkpoint calculation. In order to deal with old replicas, we change the semantics to say all *new node* in sync replicas. That means the replicas on old nodes don't count for the global checkpointing. In this state the seq# infra is not fully operational (you can't search on it, because copies may miss it) but it is maintained on shards that can support it. The old replicas will have to go through a file based recovery at some point and will get the seq# information at that point. There is still an edge case where a new primary fails and an old replica takes over. I'lll discuss this one with @ywelsch as I prefer to avoid it completely.
This PR also re-enables the BWC tests which were disabled. As such it had to fix any BWC issue that had crept in. Most notably an issue with the removal of the `timestamp` field in #21670.
The commit also includes a fix for the default value of the seq number field in replicated write requests (it was 0 but should be -2), that surface some other minor bugs which are fixed as well.
Last - I added some debugging tools like more sane node names and forcing replication request to implement a `toString`
This commit exposes public getters for the aggregations in
AggregatorFactories.Builder. The reason is that it allows to
parse the aggregation object from elsewhere (e.g. a plugin) and then
be able to get the aggregation builders in order to set them in
a SearchSourceBuilder.
The alternative would have been to expose a setter for the
AggregatorFactories.Builder object. But that would be making
the API a bit trappy.
Today if a settings object has many keys ie. if somebody specifies
a gazillion synonym in-line (arrays are keys ending with ordinals) operations like
`Settings#getByPrefix` have a linear runtime. This can cause index creations to be
very slow producing lots of garbage at the same time. Yet, `Settings#getByPrefix` is called
quite frequently by group settings etc. which can cause heavy load on the system.
While it's not recommended to have synonym lists with 25k entries in-line these use-cases should
not have such a large impact on the cluster / node. This change introduces a view-like map
that filters based on the prefixes referencing the actual source map instead of copying all values
over and over again. A benchmark that adds a single key with 25k random synonyms between 2 and 5 chars
takes 16 seconds to get the synonym prefix 200 times while the filtered view takes 4 ms for the 200 iterations.
This relates to https://discuss.elastic.co/t/200-cpu-elasticsearch-5-index-creation-very-slow-with-a-huge-synonyms-list/69052
With this commit, we introduce a cache to the geoip ingest processor.
The cache is enabled by default and caches the 1000 most recent items.
The cache size is controlled by the setting `ingest.geoip.cache_size`.
Closes#22074
In some cases, it might happen that the `_all` field gets a field type that is
not totally configured, and in particular lacks analyzers. This is due to the
fact that `AllFieldMapper.TypeParser.getDefault` uses `Defaults.FIELD_TYPE` as
a default field type, which does not have any analyzers configured since it
does not know about the default analyzers.
With this commit we enable the Jackson feature 'STRICT_DUPLICATE_DETECTION'
by default for all XContent types (not only JSON).
We have also changed the name of the system property to disable this feature
from `es.json.strict_duplicate_detection` to the now more appropriate name
`es.xcontent.strict_duplicate_detection`.
Relates elastic/elasticsearch#19614
Relates elastic/elasticsearch#22073
With this commit we change the data type of the 'TIMESTAMP'
meta-data field from a formatted date string to a plain
`java.util.Date` instance. The main reason for this change is
that our benchmarks have indicated that this contributes
significantly to the time spent in the ingest pipeline.
The overhead in terms of indexing throughput of the ingest
pipeline is about 15% and breaks down roughly as follows:
* 5% overhead caused by the conversion from `XContent` -> `Map`
* 5% overhead caused by the timestamp formatting
* 5% overhead caused by the conversion `Map` -> `XContent`
Relates #22074
In #22094 we introduce a test-only setting to simulate transport
impls that don't support handshakes. This commit implements the same logic
without a setting.
This commit touches addresses issues related to recovery and sequence numbers:
- A sequence number can be assigned and a Lucene commit created with a
maximum sequence number at least as large as that sequence number,
yet the operation corresponding to that sequence number can be
missing from both the Lucene commit and the translog. This means that
upon recovery the local checkpoint will be stuck at or below this
missing sequence number. To address this, we force the local
checkpoint to the maximum sequence number in the Lucene commit when
opening the engine. Note that there can still be gaps in the history
in the translog but we do not address those here.
- The global checkpoint is transferred to the target shard at the end
of peer recovery.
- Additionally, we reenable the relocation integration tests.
Lastly, this work uncovered some bugs in the assignment of sequence
numbers on replica operations:
- setting the sequence number on replica write requests was missing,
very likely introduced as a result of resolving merge conflicts
- handling operations that arrive out of order on a replica and have a
version conflict with a previous operation were never marked as
processed
Relates #22212
Some expert users like UnicastZenPing today establishes real connections to nodes during it's ping
phase that can be used by other parts of the system. Yet, this is potentially dangerous
and undesirable unless the nodes have been fully verified and should be connected to in the
case of a cluster state update or if we join a newly elected master. For use-cases like this, this change adds the infrastructure to manually handle connections that are not publicly available on the node ie. should not be managed by `Transport`/`TransportSerivce`
Some of our stats serialization code duplicates complicated seriazliation logic
or could use existing building blocks from StreamOutput/Input. This commit
cleans up some of the serialization code.
Today in the codebase we refer to seccomp everywhere instead of system
call filter even if we are not specifically referring to Linux. This
commit is a purely mechanical change to refer to system call filter
where appropriate instead of the general seccomp, and only leaves
seccomp in place when actually referring to the Linux implementation.
Relates #22243
We try to install a system call filter on various operating systems
(Linux, macOS, BSD, Solaris, and Windows) but the setting
(bootstrap.seccomp) to control this is named after the Linux
implementation (seccomp). This commit replaces this setting with
bootstrap.system_call_filter. For backwards compatibility reasons, we
fallback to bootstrap.seccomp and log a deprecation message if
bootstrap.seccomp is set. We intend to remove this fallback in
6.0.0. Note that now is the time to make this change it's likely that
most users are not making this setting anyway as prior to version 5.2.0
(currently unreleased) it was not necessary to configure anything to
enable a node to start up if the system call filter failed to install
(we marched on anyway) but starting in 5.2.0 it will be necessary in
this case.
Relates #22226
With recent changes to our parsing code we have drastically reduced the places where we auto-detect the content type from the input. The usage of these methods spread in our codebase for no reason, given that in most of the cases we know the content type upfront and we don't need any auto-detection mechanism. Deprecating these methods is a way to try and make sure that these methods are carefully used, and hopefully not introduced in newly written code.
We have yet to fix the REST layer to read the Content-Type header, which is the long term solution, but for now we just want to make sure that the usage of these methods doesn't spread any further.
Relates to #19388
The JSON processor has an optional field called "target_field".
If you don't specify target_field then target_field becomes what you specified as "field".
There isn't anyway to add the fields to the root of a document. By
setting `add_to_root`, now serialized fields will be inserted into the
top-level fields of the ingest document.
Closes#21898.
In #20305, _suggest endpoint was deprecated
in favour of using _search endpoint. This
commit removes the dedicated _suggest endpoint
entirely from master.
When using a bulk processor in test, you might write something like:
```java
BulkProcessor bulkProcessor = BulkProcessor.builder(client, new BulkProcessor.Listener() {
@Override public void beforeBulk(long executionId, BulkRequest request) {}
@Override public void afterBulk(long executionId, BulkRequest request, BulkResponse response) {}
@Override public void afterBulk(long executionId, BulkRequest request, Throwable failure) {}
})
.setBulkActions(10000)
.setFlushInterval(TimeValue.timeValueSeconds(10))
.build();
for (int i = 0; i < 10000; i++) {
bulkProcessor.add(new IndexRequest("foo", "bar", "doc_" + i)
.source(jsonBuilder().startObject().field("foo", "bar").endObject()
));
}
bulkProcessor.flush();
client.admin().indices().prepareRefresh("foo").get();
SearchResponse response = client.prepareSearch("foo").get();
// response does not contain any hit
```
The problem is that by default bulkProcessor defines the number of concurrent requests to 1 which is using behind the scene an Async BulkRequestHandler.
When you call `flush()` in a test, you expect it to flush all the content of the bulk so you can search for your docs.
But because of the async handling, there is a great chance that none of the documents has been indexed yet when you call the `refresh` method.
We should advice in our Java guide to explicitly set concurrent requests to `0` so users will use behind the scene the Sync BulkRequestHandler.
```java
BulkProcessor bulkProcessor = BulkProcessor.builder(client, new BulkProcessor.Listener() {
@Override public void beforeBulk(long executionId, BulkRequest request) {}
@Override public void afterBulk(long executionId, BulkRequest request, BulkResponse response) {}
@Override public void afterBulk(long executionId, BulkRequest request, Throwable failure) {}
})
.setBulkActions(5000)
.setFlushInterval(TimeValue.timeValueSeconds(10))
.setConcurrentRequests(0)
.build();
```
Closes#22158.
Depending on how the connection is closed the `#onChannelClosed` callback
might be invoked more than once or the handler has been processed by the response
of the handshake already. This commit only notifies the handler if was removed from
the pending map.
This commit fixes a silly doc bug where the field that represents the
total CPU time consumed by all tasks in the same cgroup was mistakenly
reported as "usage" instead of "usage_nanos".
Relates #21029
This commit adds a skip for the include segment file sizes REST tests on
nodes less than or equal to version 5.1.1 as the stats APIs did not
correctly account for this parameter prior to version 5.1.2.
Relates #21879
Sends the `error_trace` parameter with all requests sent by the
yaml test framework, including the doc snippet tests. This can be
overridden by settings `error_trace: false`. While this drift's
core's handling of the yaml tests from the client's slightly this
should only be a problem for tests that rely on the default value,
both of which I've fixed by setting the value explicitly.
This also escapes `\n` and `\t` in the `Stash dump on failure` so
the `stack_trace` is more readable.
Also fixes `RestUpdateSettingsAction` to not think of the `error_trace`
parameter as a setting.
Today we initialize Netty in a static initializer. We trigger this
method via static initializers from Netty-related classes, but we can
trigger this method earlier than we do to ensure that Netty is
initialized how we want it to be.
`ClusterService` is responsible of updating the cluster state on every node (as a response to an API call on the master and when non-masters receive a new state from the master). When a new cluster state is processed, it is made visible via the `ClusterService#state` method and is sent to series of listeners. Those listeners come in two flavours - one is to change the state of the node in response to the new cluster state (call these cluster state appliers), the other is to start a secondary process. Examples for the later include an indexing operation waiting for a shard to be started or a master node action waiting for a master to be elected.
The fact that we expose the state before applying it means that samplers of the cluster state had to worry about two things - working based on a stale CS and working based on a future, i.e., "being applied" CS. The `ClusterStateStatus` was used to allow distinguishing between the two. Working with a stale cluster state is not avoidable. How this PR changes things to make sure consumers don't need to worry about future CS, removing the need for the status and simplifying the waiting logic.
This change does come with a price as "cluster state appliers" can't sample the cluster state from `ClusterService` whenever they want as the cluster state isn't exposed yet. However, recent clean ups made this is situation easier and this PR takes the last steps to remove such sampling. This also helps clarify the "information flow" and helps component separation (and thus potential unit testing). It also adds an assertion that will trigger if the cluster state is sampled by such listeners.
Note that there are still many "appliers" that could be made a simpler, unrestricted "listener" but this can be done in smaller bits in the future. The commit also makes it clear what the `appliers` and what the `listeners` are by using dedicated interfaces.
Also, since I had to change the listener types I went ahead and changed the data structure for temporary/timeout listeners (used for the observer) so addition and removal won't be an O(n) operation.
This commit addresses an issue in the stats APIs where
include_segment_file_sizes was not being consumed leading to requests
containing this parameter being rejected.
Relates #21879
Since #22094 has been back-ported to 5.2 we can remove all BWC layers from master since all supported version will handle handshake requests.
Relates to #22094
Today sending a message on a closed channel doesn't throw an exception. The channel
might just swallow the exception and informs the internal async exception handler
that a channel got disconnected. This change adds a safety check that we fail
the handshake if we registered a handler but the channel has been closed already
for instance due to a reset by peer.
Merging mappings ensures that fields are used consistently across mapping types. Disabling norms for a specific field in one mapping type for example also disables norms for the same field in other mapping types of that index. The logic that ensures this while merging mappings currently always creates a fresh document mapper for all existing mapping types, even if no change occurred. Creating such a fresh document mapper does not come for free though as it involves recompressing the source. Making a mapping change to one type of an index with 100 types will thus re-serialize and recompress all 100 types, independent of any changes made to those types.
This commit fixes the update logic to only create a new DocumentMapper if a field type actually changes.
* Replace _suggest endpoint to _search in docs
In 5.0, the _suggest endpoint is just sugar for _search
with suggestions specified. Users should move away from
using the _suggest endpoint, as it is marked as deprecated in 5.x and
will be removed in 6.0
* update docs to use _search endpoint instead of _suggest
* Add deprecation logging to RestSuggestAction
* Use search endpoint instead of suggest endpoint in rest tests
Inline scripts defined in Ingest Pipelines are now compiled at creation time to preemptively catch errors on initialization of the pipeline.
Fixes#21842.
* Build: Remove hardcoded reference to x-plugins in build
The gradle build currenlty allows extra plugins to be hooked into the
elasticsearch build by placing under an x-plugins directory as a sibling
of the elasticsearch checkout. This change converts this directory to
one called elasticsearch-extra. The subdirectories of
elasticsearch-extra become first level projects in gradle (while before
they would have been subprojects of `:x-plugins`. Additionally, this
allows major version checkouts to be associated with extra plugins. For
example, you could have elasticsearch checked out as
`elasticsearch-5.x`, and a sibling `elasticsearch-5.x-extra` directory.
Low level handshake code doesn't handle situations gracefully if the connection
is concurrently closed or reset by peer. This commit adds the relevant code to
fail the handshake if the connection is closed.
Moves the last of the "easy" parser construction into
`RestRequest`, this time with a new method
`RestRequest#contentParser`. The rest of the production
code that builds `XContentParser` isn't "easy" because it is
exposed in the Transport Client API (a Builder) object.
Boaz Leskes
Dimitris Athanasiou
Simon Willnauer
Clinton Gormley
Daniel Mitterdorfer
Daniel Mitterdorfer
Adrien Grand
Jason Tedor
David Pilato
Ryan Ernst
Luca Cavanna
Tal Levy
Areek Zillur
Pablo Musa
Simon Willnauer
Florian Hopf
Masaru Hasegawa
Nik Everett
Tanguy Leroux
Aaron Spiegel
Jim Ferenczi
Yannick Welsch