The distribution of watches now happens on the node which holds the
watches index, instead of on the master node. This requires several
changes to the current implementation.
1. Running on shards and replicas
In order to run watches on the nodes with the watches index on its
primaries and replicas. To ensure that watches do not run twice, there is
a logic which checks the local shards, runs a murmurhash on the id and
runs modulo against the number of shards and replicas, this is the way to
find out, if a watch should run local. Reloading happens
2. Several master node actions moved to a HandledTransportAction, as they
are basically just aliases for indexing actions, among them the
put/delete/get watch actions, the acknowledgement action, the de/activate
actions
3. Stats action moved to a broadcast node action, because we potentially
have to query every node to get watcher statistics
4. Starting/Stopping watcher now is a master node action, which updates
the cluster state and then listeners acts on those. Because of this watches
can be running on two systems, if you those have different cluster state
versions, until the new watcher state is propagated
5. Watcher is started on all nodes now. With the exception of the ticker
schedule engine most classes do not need a lot of resources while running.
However they have to run, because of the execute watch API, which can hit
any node - it does not make sense to find the right shard for this watch
and only then execute (as this also has to work with a watch, that has not
been stored before)
6. By using a indexing operation listener, each storing of a watch now
parses the watch first and only stores on successful parsing
7. Execute watch API now uses the watcher threadpool for execution
8. Getting the number of watches for the stats now simply queries the
different execution engines, how many watches are scheduled, so this is
not doing a search anymore
There will be follow up commits on this one, mainly to ensure BWC compatibility.
Original commit: elastic/x-pack-elasticsearch@0adb46e658
The yaml test runner now throws error when skip or do sections are malformed, such as they don't start with the proper token (START_OBJECT). That signals bad indentation, which was previously ignored. Thanks (or due to) our pull parsing code, we were still able to properly parse the sections, yet other runners weren't able to.
Original commit: elastic/x-pack-elasticsearch@920201207c
Transform and condition scripts were only compiled on its initial creation, so
when a new watch is created or when the master node loads all the watches. However
changing a script (like a stored one) did not lead to any changes in the in memory
watch store and thus the old script was executed again.
We do however have a mechanism in Elasticsearch's ScriptService that already does some
caching, and should reuse that one.
Closeselastic/elasticsearch#4237
Original commit: elastic/x-pack-elasticsearch@477548e237
When adding a watch which has a painless component, the scriptexception
was wrapped into a deprecated exception which means, that the awesome
painless descriptions were lost. This wrapping has been removed.
Closeselastic/elasticsearch#3161
Original commit: elastic/x-pack-elasticsearch@1703fe4eb6
Alexander Reelsen
Ryan Ernst
Luca Cavanna
Jack Conradson
Nik Everett