Roman Puchkovskiy created IGNITE-18612:
------------------------------------------
Summary: Make RAFT snapshot streaming resistant to network glitches
Key: IGNITE-18612
URL: https://issues.apache.org/jira/browse/IGNITE-18612
Project: Ignite
Issue Type: Improvement
Reporter: Roman Puchkovskiy
Assignee: Roman Puchkovskiy
Fix For: 3.0.0-beta2
Network is inherently unreliable (see IGNITE-18605). RAFT snapshot streaming
might take dozens of minutes. Current implementation breaks on the first
lost/improperly handled message, so If we get a glitch in the middle of a long
snapshot installation, we'll waste a lot of resources.
The idea is to:
# Make requests idempotent (so that we can repeat a lost request).
{{SnapshotMetaRequest}} is already idempotent. {{SnapshotMvDataRequest}} and
{{SnapshotTxDataRequest}} can be provided with a sequenceNo (increased at the
receiving side for each next request, but not increased for a retry of a
previous request). As there is only one receiver, and it always works
sequentially, on the sending side we'll have to only additionally remember the
previous response and its sequenceNo to support idempotent retries that do not
cause excessive cursor advancement.
# On the receiving side, specify a sane timeout (like a few seconds) per
request and retry requests that error or timeout (using the correct sequenceNo)
# If an error happens while processing a request on the sending side, return
an indication of an error to the receiver instead of just dropping the message
(so that the receiver gets informed about the necessity to make a retry
faster, AND the receiver can see whether it should stop retrying if the failure
is fatal).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
