NRG (2.11): Completeness/consistency after leader changes #6485
+182
−72
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
| if n.pindex > n.applied { | ||
| n.aflr = n.pindex | ||
| } else { | ||
| n.aflr = 0 |
There was a problem hiding this comment.
Not entirely clear what 0 means here, I know the comment above says signalling disabled, but is that implying this doesn't track upwards with applied always?
There was a problem hiding this comment.
We only need to track if we have a log with entries that are not applied. If all our entries are applied we can signal immediately. Resetting n.aflr is just for sanity so we couldn't signal leader twice.
Have added this comment:
// We know we have applied all entries in our log and can signal immediately.
// For sanity reset applied floor back down to 0, so we aren't able to signal twice.
Signed-off-by: Maurice van Veen <github@mauricevanveen.com>
MauriceVanVeen
force-pushed
the
maurice/nrg-leader-change-consistency
branch
from
8916405 to
43e7a02
Compare
derekcollison
added a commit
that referenced
this pull request
Feb 10, 2025
Reverts #6415 and #6426. If duplication would be deferred to be done by a replica, and that replica was down for at least 'dupe window+5s', and it would clean up the dedupe map. A message that was meant to be duplicate could be passed as a genuine message, resulting in stream desync. Instead, the dedupe map should be cleared of any staged zero-sequences. However, that was not possible before as a new leader would not always be fully up-to-date when it starts responding to new write requests, which could result in duplicate messages as well. However, relying on the 'Leader Completeness Property' implemented here: #6485, we can confidently clear the dedupe map now of any staged zero-sequences (knowing they were not proposed). Ensuring both there's no desync, and a failed proposal for a message would not block subsequent messages with the same dedupe ID. Have left the commits as separate, to ease reviewing. Signed-off-by: Maurice van Veen <github@mauricevanveen.com>
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.
Previously #6194 implemented a way to wait for entries that are stored in the new leader's log but not yet applied, before allowing 'expected per subject' operations to go through. This protected against KV/stream desync.
That protection is now extended to the whole of (clustered) JetStream, and not specific to this one operation anymore. Meaning that if the new leader recognizes it has entries in its log that have not yet been applied, it waits for those to be applied before responding to any read/write operations. In essence it's the 'Leader Completeness Property' as described by the Raft paper. This also brings us closer to 'read-your-writes' when only requesting reads from the leader.
Signed-off-by: Maurice van Veen github@mauricevanveen.com