[CCR] Made shard follow task more resilient against node failure and #31242
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
martijnvg
added
WIP
:Distributed/CCR
|
Pinging @elastic/es-distributed |
|
I've marked this PR as WIP, because the test added in this pr sometimes fails with a assertion like this one: |
|
The reason that this assertion trips is because the primary term for the two translog operations with the same seqno are different. With normal indexing / replication this error should not happen, but with ccr this can currently happen. In this case particular situation, the node with a follow index primary was closed while the shard follow task was indexing into this shard. The shard follow task then re-sends the same translog operations to the promoted replica shard copy and new replica shard copy. At the same time this new replica shard is also doing shard recovery. This causes the assertion to trip, because the two translog operations with same data and seqno number, but with a different primary term are detected. Besides this assertion tripping, the follower shard now also contains an additional document that the leader does not have. This document for the write operation with the lowest primary term is soft deleted, but if this follow shard becomes a leader shard for another follow index setup then we will also replicate this surprise write operation the second follow index. Outside of this assertion failure the test that closes a node while follow an index test passes and the changes to |
added a test that verifies we can close a node while following an index.
martijnvg
force-pushed
the
ccr_close_node_test
branch
from
82e13bd
to
48e7371
Compare
|
Superseded by: #31581 |
martijnvg
added a commit
to martijnvg/elasticsearch
that referenced
this pull request
Jun 27, 2018
The current shard follow mechanism is complex and does not give us easy ways the have visibility into the system (e.g. why we are falling behind). The main reason why it is complex is because the current design is highly asynchronous. Also in the current model it is hard to apply backpressure other than reducing the concurrent reads from the leader shard. This PR has the following changes: * Rewrote the shard follow task to coordinate the shard follow mechanism between a leader and follow shard in a single threaded manner. This allows for better unit testing and makes it easier to add stats. * All write operations read from the shard changes api should be added to a buffer instead of directly sending it to the bulk shard operations api. This allows to apply backpressure. In this PR there is a limit that controls how many write ops are allowed in the buffer after which no new reads will be performed until the number of ops is below that limit. * The shard changes api includes the current global checkpoint on the leader shard copy. This allows reading to be a more self sufficient process; instead of relying on a background thread to fetch the leader shard's global checkpoint. * Reading write operations from the leader shard (via shard changes api) is a seperate step then writing the write operations (via bulk shards operations api). Whereas before a read would immediately result into a write. * The bulk shard operations api returns the local checkpoint on the follow primary shard, to keep the shard follow task up to date with what has been written. * Moved the shard follow logic that was previously in ShardFollowTasksExecutor to ShardFollowNodeTask. * Moved over the changes from elastic#31242 to make shard follow mechanism resilient from node and shard failures. Relates to elastic#30086
martijnvg
added a commit
that referenced
this pull request
Jul 10, 2018
The current shard follow mechanism is complex and does not give us easy ways the have visibility into the system (e.g. why we are falling behind). The main reason why it is complex is because the current design is highly asynchronous. Also in the current model it is hard to apply backpressure other than reducing the concurrent reads from the leader shard. This PR has the following changes: * Rewrote the shard follow task to coordinate the shard follow mechanism between a leader and follow shard in a single threaded manner. This allows for better unit testing and makes it easier to add stats. * All write operations read from the shard changes api should be added to a buffer instead of directly sending it to the bulk shard operations api. This allows to apply backpressure. In this PR there is a limit that controls how many write ops are allowed in the buffer after which no new reads will be performed until the number of ops is below that limit. * The shard changes api includes the current global checkpoint on the leader shard copy. This allows reading to be a more self sufficient process; instead of relying on a background thread to fetch the leader shard's global checkpoint. * Reading write operations from the leader shard (via shard changes api) is a separate step then writing the write operations (via bulk shards operations api). Whereas before a read would immediately result into a write. * The bulk shard operations api returns the local checkpoint on the follow primary shard, to keep the shard follow task up to date with what has been written. * Moved the shard follow logic that was previously in ShardFollowTasksExecutor to ShardFollowNodeTask. * Moved over the changes from #31242 to make shard follow mechanism resilient from node and shard failures. Relates to #30086
martijnvg
added a commit
that referenced
this pull request
Jul 10, 2018
The current shard follow mechanism is complex and does not give us easy ways the have visibility into the system (e.g. why we are falling behind). The main reason why it is complex is because the current design is highly asynchronous. Also in the current model it is hard to apply backpressure other than reducing the concurrent reads from the leader shard. This PR has the following changes: * Rewrote the shard follow task to coordinate the shard follow mechanism between a leader and follow shard in a single threaded manner. This allows for better unit testing and makes it easier to add stats. * All write operations read from the shard changes api should be added to a buffer instead of directly sending it to the bulk shard operations api. This allows to apply backpressure. In this PR there is a limit that controls how many write ops are allowed in the buffer after which no new reads will be performed until the number of ops is below that limit. * The shard changes api includes the current global checkpoint on the leader shard copy. This allows reading to be a more self sufficient process; instead of relying on a background thread to fetch the leader shard's global checkpoint. * Reading write operations from the leader shard (via shard changes api) is a separate step then writing the write operations (via bulk shards operations api). Whereas before a read would immediately result into a write. * The bulk shard operations api returns the local checkpoint on the follow primary shard, to keep the shard follow task up to date with what has been written. * Moved the shard follow logic that was previously in ShardFollowTasksExecutor to ShardFollowNodeTask. * Moved over the changes from #31242 to make shard follow mechanism resilient from node and shard failures. Relates to #30086
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
added a test that verifies we can close a node while following an index.