Hi Alan,
To add to Daniel’s response, as part of
https://issues.apache.org/jira/browse/HIVE-18264 and
https://issues.apache.org/jira/browse/HIVE-18661 (I’m actively working on
these), we plan to remove the current mechanism of updating the cache (which is
very inefficient anyway) and instead use the NOTIFICATION_LOG table to update
the cache incrementally. The code that you pointed was meant to not let the
background update thread block the metastore client calls for a long time, but
with the plan to update the cache incrementally we may not need to worry about
that, as applying the notification incrementally will not be a long blocking
execution.
Thanks,
--Vaibhav
On 2/8/18, 11:41 AM, "Daniel Dai" wrote:
Hi, Alan,
If database cache is changed locally, we don’t want to bring remote copy to
overwrite it as the remote copy doesn’t carry local changes (ideally, we shall
also apply local changes to the remote copy images we bring in from db, but we
are not there yet). That’s why we skip the update if there’s local changes, and
wait for the next iteration to sync with remote. isDatabaseCacheDirty is
initially set to false unless there’s local update, and will be reset during
cache swap, thus give a chance for the next iteration to update the cache if
there’s no local changes.
Thanks,
Daniel
On 2/6/18, 11:57 AM, "Alan Gates" wrote:
I’m confused by the following code in the CachedStore. This in in the
CacheUpdateMasterWork thread, in the updateDatabases method (which is
called by update()):
*// Skip background updates if we detect change*
*if *(*isDatabaseCacheDirty*.compareAndSet(*true*, *false*)) {
*LOG*.debug(*"Skipping database cache update; the database list we
have
is dirty."*);
*return*;
}
Why are we not updating the cache if we’ve dirtied it? Also, AFAICT no
one
ever sets isDatabaseCacheDirty to false, meaning once one database is
created the cache will never be updated. Am I missing something?
Alan.