Blake Eggleston created CASSANDRA-21791:
-------------------------------------------
Summary: CEP-45 - improve tracked read coordination
Key: CASSANDRA-21791
URL: https://issues.apache.org/jira/browse/CASSANDRA-21791
Project: Apache Cassandra
Issue Type: Bug
Reporter: Blake Eggleston
The tracked read coordinator currently waits for all peers to send it both
their initial summaries and their acknowledgement that they’ve received any
mutations they were missing. This is the network equivalent of running read
repair for every read. This patch improves read completion by:
* Not waiting for sync acks from nodes we know are up to date. We know which
writes each peer is missing from their initial summary, if we can see that
they’re not behind, we no longer wait for them to send sync acknowledgements.
* Sending per-peer sync acks. Peers currently wait to receive all mutations
they’re missing from all other peers before notifying the coordinator. This
send an ack per-peer
* Accounting for mutations received between local summary creation and remove
summary receipt
* Eager quorum completion - both of these improvements allow us to be more
intelligent about how we decide we’ve reached quorum. If a quorum exists that
we can satisfy immediately (for instance, 1/3 non data nodes is reporting a
mutation none of the other peers have seen, we can skip syncing and returning
that mutation while preserving monotonicity)
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
