In order to safely complete recoveries / relocations we have to keep all operation done since the recovery start at available for replay. At the moment we do so by preventing the engine from flushing and thus making sure that the operations are kept in the translog. A side effect of this is that the translog keeps on growing until the recovery is done. This is not a problem as we do need these operations but if the another recovery starts concurrently it may have an unneededly long translog to replay. Also, if we shutdown the engine for some reason at this point (like when a node is restarted) we have to recover a long translog when we come back.
To void this, the translog is changed to be based on multiple files instead of a single one. This allows recoveries to keep hold to the files they need while allowing the engine to flush and do a lucene commit (which will create a new translog files bellow the hood).
Change highlights:
- Refactor Translog file management to allow for multiple files.
- Translog maintains a list of referenced files, both by outstanding recoveries and files containing operations not yet committed to Lucene.
- A new Translog.View concept is introduced, allowing recoveries to get a reference to all currently uncommitted translog files plus all future translog files created until the view is closed. They can use this view to iterate over operations.
- Recovery phase3 is removed. That phase was replaying operations while preventing new writes to the engine. This is unneeded as standard indexing also send all operations from the start of the recovery to the recovering shard. Replay all ops in the view acquired in recovery start is enough to guarantee no operation is lost.
- IndexShard now creates the translog together with the engine. The translog is closed by the engine on close. ShadowIndexShards do not open the translog.
- Moved the ownership of translog fsyncing to the translog it self, changing the responsible setting to `index.translog.sync_interval` (was `index.gateway.local.sync`)
Closes#10624
Currently we use 5k operations as a flush threshold. Indexing 5k documents
per second is rather common which would cause the index to be committed on
the lucene level each time the flush logic runs which is 5 seconds by default.
We should rather use a size based threshold similar to the lucene index writer
that doesn't cause such agressive commits which can slow down indexing significantly
especially since they cause the underlying devices to fsync their data.
Clinton Gormley
Boaz Leskes
Michael McCandless
Simon Willnauer
Shay Banon