Boyi Dai created SOLR-18562:
-------------------------------
Summary: Fixing a time routed alias's date math start can remove
another alias from aliases.json
Key: SOLR-18562
URL: https://issues.apache.org/jira/browse/SOLR-18562
Project: Solr
Issue Type: Bug
Components: SolrCloud
Reporter: Boyi Dai
When a time routed alias is created with date math in router.start (e.g.
NOW/DAY), the first document routed through it fixes the start to an instant
(SOLR-13760). TimeRoutedAlias.validateRouteValue() takes a snapshot of the
node's cached aliases, adds the new router.start to the alias's properties, and
then writes it asynchronously with
AliasesManager.applyModificationAndExportToZk().
The modification function ignores the current aliases it is passed and returns
the snapshot instead:
{code:java}
Aliases aliases = zkStateReader.getAliases();
...
core.runAsync(
() ->
zkStateReader.aliasesManager.applyModificationAndExportToZk(
(a) -> aliases.cloneWithCollectionAliasProperties(aliasName,
props)));
{code}
applyModificationAndExportToZk() writes with the znode version of the current
aliases and, on a BadVersionException, refreshes and calls the function again.
Since the function always returns the old snapshot, the retry writes the old
collection aliases over the new version. Any alias created (or alias property
changed) after the snapshot was taken is silently removed from aliases.json.
Every other caller of applyModificationAndExportToZk builds on its argument.
This was found through code analysis.
*Steps to reproduce:*
# Create a time routed alias A with router.start=NOW/DAY.
# Create another alias B before the node that will receive A's first document
has seen it (or before the asynchronous write runs).
# Send the first document for A to that node.
# Result: router.start is fixed, but B is gone from aliases.json. Requests to B
fail, and refreshing doesn't bring it back since the alias is lost in ZooKeeper.
*Expected:* only A's router.start changes; other aliases are untouched.
*Proposed fix:* apply the change to the aliases passed to the modification
function, and only set router.start instead of copying all of A's properties
from the snapshot. I have a patch with a test and will open a PR.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
