Nándor Kollár created IMPALA-15605:
--------------------------------------
Summary: REST catalog DML deletes committed files when the commit
response is lost
Key: IMPALA-15605
URL: https://issues.apache.org/jira/browse/IMPALA-15605
Project: IMPALA
Issue Type: Sub-task
Reporter: Nándor Kollár
A network outage is not treated as commit-state-unknown. For a refused
connection that's fine, because the request never arrived. But in the worst
case:
1. the request reaches the REST server, and the server commits the new
snapshot;
2. the connection drops before the response arrives (a read timeout or a
reset);
3. Impala gets a RESTException, treats it as a failed commit, and deletes the
data, delete and Puffin files that the committed snapshot references.
This sequence might result in a corrupted table.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
