As there are two indices to upgrade for watcher, it makes a lot of sense
to also have two upgrade checks.
There is one upgrader for the watches index, which deletes
old templates, adds the new one before and then does the reindexing.
Same for the triggered watches index.
This also means, that there will be two entries popping up in the kibana
UI.
Note: Each upgrade check checks if the other index (for the .watches
upgrade check the triggered watches index and vice versa) is already
upgraded and only if that is true, watcher is restarted.
relates elastic/x-pack-elasticsearch#2238
Original commit: elastic/x-pack-elasticsearch@2c92040ed6
I noticed this while working on a previous issue with atomic move writer
(silent swallowing of exceptions). Namely, atomic move writer has
dangerous semantics. The problem is as follows: atomic move writer works
by writing lines to a temporary file, and then in its close method it
replaces the target path with the temporary file. However, the close
method is invoked whether or not all writes to the temporary file
succeeded (because writers obtained from atomic move writer are used in
try-with-resources blocks, as they should be). There is no way to
distinguish that the writer is being closed in a successful scenario
versus a failure scenario. In the close method for atomic move writer,
the target file is replaced by the temporary file. This means that the
target file is replaced whether or not writing to the temporary file
actually succeeded. Since these atomic move writers are used for user
configuration files (users and user_roles), a failure here can lead to
data loss for the user, a tragedy!
There is another (less serious) problem with the atomic move
writer. Since the close method tries to move the temporary file in place
of the existing file, the temporary file can be left behind if there is
another failure in the close method (e.g., closing the underlying file
after writing, or setting the permissions on the temporary file). This
means that in some situations, atomic move writer will leave temporary
files behind (which is not definitively not atomic).
This commit replaces the atomic move writer with a safer mechanism. We
still perform the write atomically in the following sense: we write to a
temporary file. Either writing to that file succeeds or it fails. If
writing succeeds, we replace the existing file with the temporary
file. If writing fails, we clean up the temporary file and the existing
file remains in place.
Relates elastic/x-pack-elasticsearch#2299
Original commit: elastic/x-pack-elasticsearch@3199decb0a
Certain types of datafeeds cannot have null chunking configs, so
setting chunking config to null sometimes doesn't stick as null
Original commit: elastic/x-pack-elasticsearch@3a52bad460
When a watch is executed currently, it gets passed an in-memory
watch object, that was loaded, before the execution started.
This means there is a window of time, where an old watch could still
be executing, then a watch gets loaded for execution, then the old watch
execution finishes and updates the watch status and thus reindexes the
watch.
Now the watch, that got loaded for execution, executes and tries to
store its watch status, but fails, because the version of the watch
has changed.
This commit changes the point in time where the watch is loaded. Now
this only happens, while a watch is in its protected execution block,
and thus we can be sure, that there is no other execution of the watch
happening.
This will primarily impact watches, that execute often, but their
runtime is longer than the configured interval between executions.
Side fix: Removed some duplicate testing method and moved into
WatcherTestUtils, fixed a tests with a ton of if's with random booleans
into separate tests.
relates elastic/x-pack-elasticsearch#395
Original commit: elastic/x-pack-elasticsearch@bf393023d7
Today it's impossible to run xpack bwc tests against any other branch
or remote than upstream. This allows to pass `-Dtests.bwc.refspec` to
change the refspec to use for the bwc tests.
Original commit: elastic/x-pack-elasticsearch@4d365f5a6e
Today we require a pre-shared key to use the token service. Beside the
additional setup step it doesn't allow for key-rotation which is a major downside.
This change adds a TokenService private ClusterState.Custom that is used to distribute
the keys used to encrypt tokens. It also has the infrastructur to add automatic key
rotation which is not in use yet but included here to illustrate how it can work down
the road.
This is considered a prototype and requires additioanl integration testing. Yet, it's fully
BWC with a rolling / full cluster restart from a previous version (also from 5.6 to 6.x)
since if the password is set it will just use it instead of generating a new one.
Once we implement the automatic key rotation via the clusterstate we need to ensure that we are
fully upgraded before we do that.
Also note that the ClusterState.Custom is fully transient and will never be serialized to disk.
Original commit: elastic/x-pack-elasticsearch@1ae22f5d41
This change means that newly created jobs will get an explicit 1GB
model memory limit if no model memory limit is specified when creating
the job. Existing jobs that had a null model memory limit will carry
on using the default model memory limit defined in the C++ code.
Relates elastic/x-pack-elasticsearch#546
Original commit: elastic/x-pack-elasticsearch@a4e6b73c2b
We rely on command extensions in our scripts but we do not actually
guarantee that they are enabled (usually they are, by default, but they
can be disabled outside of our control). This commit ensures that they
are enabled.
Relates elastic/x-pack-elasticsearch#2307
Original commit: elastic/x-pack-elasticsearch@a5eec8ca7b
Today we require the `bootstrap.password` to be present in the keystore in order to
bootstrap xpack. With the addition of `keystore.seed` we have a randomly generated password
per node to do the bootstrapping. This will improve the initial user experience significantly
since the user doesn't need to create a keystore and add a password, they keystore is created
automatically unless already present and is always created with this random seed.
Relates to elastic/elasticsearch#26253
Original commit: elastic/x-pack-elasticsearch@5a984b4fd8
The old message of "Cannot auto close job" implied the problem was with
closing the job. This change makes it clearer that the problem is that
the datafeed could not be stopped and hence auto-close will not even be
attempted.
Original commit: elastic/x-pack-elasticsearch@065e9930ce
These members are default initialized on contruction and then set by the
init() method. It's possible that another thread accessing the object
after init() is called could still see the null/0 values, depending on how
the compiler optimizes the code.
Original commit: elastic/x-pack-elasticsearch@668121e274
Today we try to bootstrap the security index with the bootstrap password and recommend the user to change the password with the user tool. This is trappy for instance if you happen to configure multiple nodes with a different bootstrap passwords (which is possible) it's unclear which password made it too bootstrap. Yet, we tell in the logs but it can still be very confusing. In general it should be possible to bootstrap with the user tool from any node unless the user is already created in the native user store. This change uses the bootstrap.password from the local node and always authenticate against it until the user is bootstrapped even if the passwords are different on different nodes. This will also work for authenticating against the cluster for instance if a user deletes the .security index or if that index has not been upgraded.
Original commit: elastic/x-pack-elasticsearch@8cebecb287
When writing the users and users_roles files, we wrap a custom writer in
a print writer. There is a problem with this though: when print writer
closes it closes our underlying custom writer and the close
implementation for our custom writer is not trivial, it executes code
that can throw an I/O exception. When print writer invokes this close
and an I/O exception is thrown, it swallows that exception and sets the
status on the print writer to error. One would think that we could
simply check this status but alas print writer is broken here. The act
of checking the status causes print writer to try to flush the
underyling stream which is going to be completely undefined because the
underlying stream might or might not be closed. This might cause another
exception to be thrown, losing the original. Print writer screwed the
pooch here, there is no good reason to try to do any I/O after the
underlying writer entered a failed state. To address this we remove the
use of print writer, we use our custom writer directly. This allows any
thrown exceptions to bubble up.
Relates elastic/x-pack-elasticsearch#2288
Original commit: elastic/x-pack-elasticsearch@11b8dd5641
When mappings are updated for an index are updated most settings are
merged, but not _meta. This change ensures that _meta is set when we
add per-job term mappings to our results index mappings. In order to
keep the logic for updating mappings after upgrade working, we now
have to put ALL the mappings for our results along with the latest _meta
section when updating per-job term mappings.
relates elastic/x-pack-elasticsearch#2265
Original commit: elastic/x-pack-elasticsearch@f58c11a13e
We close the secure settings in core before we pull bootstrap checks.
This means if a bootstrap check like the `TokenPassphraseBootstrapCheck`
accesses a secure setting that late it will fail due to an exception in
the `PKCS12KeyStore`. This change moves the bootstrap check creation
to the plugin constructor and adds a dummy setting to the integTest
that triggers the bootstrap checks.
Original commit: elastic/x-pack-elasticsearch@2b20865d1c
When the machine-learning-cpp repo is built locally, the zip file it
creates is preferred over that downloaded from s3 when creating the
overall x-pack-elasticsearch zip. However, prior to this change the
build would ALSO download an ml-cpp zip from s3, and just not use it.
Original commit: elastic/x-pack-elasticsearch@bd71637edd
PUT /_xpack/license with no content or content-type should fail with an appropriate error message rather than throwing NPE.
Original commit: elastic/x-pack-elasticsearch@f8c744d2a2
This change makes 2 improvements to the max_running_jobs setting:
1. Namespaces it by adding the xpack.ml. prefix
2. Renames "running" to "open", because the "running" terminology
is not used elsewhere
The old max_running_jobs setting is used as a fallback if the new
xpack.ml.max_open_jobs setting is not specified. max_running_jobs
is deprecated and (to ease backporting in the short term) will be
removed from 7.0 in a different PR closer to release of 7.0.
Relates elastic/x-pack-elasticsearch#2185
Original commit: elastic/x-pack-elasticsearch@18c539f9bb
These tests used to fail rarely, because during a watch execution
one of the watcher shards was relocated resulting in a second execution
of watch.
In order to prevent this, the tests do not need to actually create any
shards, which causes watcher potentially to be rebalanced.
This simplifies and speeds up the test as well.
relates elastic/x-pack-elasticsearch#1608
Original commit: elastic/x-pack-elasticsearch@1cfac1145d
This cleans up logging, when starting several elasticsearch instances,
as otherwise you cannot see, which node emits this log message.
Original commit: elastic/x-pack-elasticsearch@c8c2819d86
When a watch is executed, it sends an update request to the watch to
udpate its status.
This update request also updates the status.state field, which contains
information, if the watch is active. If the watch gets executed, and
during execution a watch gets disabled, then the current execution will
set the watch back to active.
This commit fixes the current behaviour and never changes the state of
a watch when updating the status after executing, allowing
activate/deactivate calls to work as expected, regardless if a watch
is being executed.
This will fix not only the current behaviour but also some flaky tests.
Original commit: elastic/x-pack-elasticsearch@ca69109ecb
It is really hard to debug some issues with watcher, when only the
e.getMessage() is returned as failure reasons instead of the whole
stack trace.
This commit gets rid of ExceptionsHelper.detailedMessage(e) and always
returns the whole exception.
This commit also extends the watch history to have all fields named
error be treated like an object to be sure they do not get
indexed. No matter where it's placed in the hierarchy
In addition a few Field interface classes were removed, that only contained parse fields.
relates elastic/x-pack-elasticsearch#1816
Original commit: elastic/x-pack-elasticsearch@b2ce680139
This commit adds the max_running_jobs setting from elasticsearch.yml
into a node attribute called ml.max_open_jobs. Previously there was
an assumption that max_running_jobs would be the same for all nodes in
the cluster. However, during a rolling cluster restart where the value
of the setting is being changed this clearly cannot be the case, and
would cause unexpected/unpredictable limits to be used during the period
when different nodes had different settings.
For backwards compatibility, if another node in the cluster has not added
its setting for max_running_jobs to the cluster state then the old
(flawed but better than nothing) approach is applied, i.e. assume the
remote node's setting for max_running_jobs is equal to that of the node
deciding the job allocation.
Relates elastic/x-pack-elasticsearch#2185
Original commit: elastic/x-pack-elasticsearch@1e62b89183
Validating job groups during parsing results into
the validation error being wrapped into a parse
exception. The UI then does not display the cause of the
error. Finally, it is conceptually not a parse error, so
it belongs outside the parsing phase.
Original commit: elastic/x-pack-elasticsearch@a03f002bdc
Only unit tests were broken. Production ML code was always terminating
bulk requests with newlines.
Original commit: elastic/x-pack-elasticsearch@96ed06fed3
If one of the old watcher templates does not exist when we try
to delete it, the upgrade should just continue.
Original commit: elastic/x-pack-elasticsearch@6a52bad329
This removes the `IndicesStatsCollector` and, instead, it reuses the superset version of the call from the `IndexStatsCollector`.
On clusters with a large number of indices, this should actually help a good amount in reducing wasted calls and memory allocation without any difference in the output.
Original commit: elastic/x-pack-elasticsearch@93b09878e4
This commit re-enables the OpenLDAP tests that were previously running against a one-off instance
in AWS but now run against a vagrant fixture. There were some IntegTests that would run against the
OpenLDAP instance randomly but with this change they no longer run against OpenLDAP. This is ok as
the functionality that is tested by these has coverage elsewhere.
relates elastic/x-pack-elasticsearch#1823
Original commit: elastic/x-pack-elasticsearch@ac9bc82297
record_count is no longer written to new results, but is still tolerated
for backwards compatibility. However, in the backwards compatibility case
the results index must already contain the required mapping. There's no
need to add this mapping to newly created results indices.
Original commit: elastic/x-pack-elasticsearch@e586f3ba96
Fix TemplateTransformMappingTests to work, even if date rolls over
during execution.
Reenable test in BootStrapTests, was forgotten.
Remove the SecurityF/MonitoringF/WatcherF classes, as there is a gradle
command to easily start elasticsearch with xpack
Remove HasherBenchmark, as it is not a test and relies on RandomContext
that is not available anymore (also I think a JMH benchmark would be
needed here).
Remove ManualPublicSmtpServersTester, was not usable anymore.
Remove OldWatcherIndicesBackwardsCompatibilityTests, now in dedicated
rolling upgrade tests.
Remove unused EvalCron class.
Original commit: elastic/x-pack-elasticsearch@100fa9e9b0
Alexander Reelsen
Jason Tedor
David Roberts
Colin Goodheart-Smithe
Simon Willnauer
Yannick Welsch
Tim Vernum
Dimitris Athanasiou
Chris Earle
Tim Sullivan
Jay Modi
David Kyle