Suppose that:
Each call to getConnection() will trigger a call to getReaderIndex() on group "a", which will most likely fail and the DBConnRef will use a "generic" server; in rare cases, it might succeed and use a group "b" server (with different REPEATABLE-READ view snapshot than the group "a" server connection). This is inefficient and might add log spam.
| Subject | Repo | Branch | Lines +/- | |
|---|---|---|---|---|
| rdbms: simplify query group selection in LoadBalancer::getConnection() | mediawiki/core | master | +18 -17 |
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Resolved | aaron | T326187 Make LoadBalancer::getReaderIndex handle down groups better | |||
| Resolved | Ladsgroup | T326274 Reduce complexity of LB and LBF (introducing IConnectionProvider) | |||
| Resolved | Ladsgroup | T325389 Deprecate ILoadBalancer::getAnyOpenConnection() | |||
| Resolved | Ladsgroup | T330590 External LBs should not be exposed to developers | |||
| Resolved | Ladsgroup | T343917 Support ReadOnlyMode without internal LoadBalancer |
A simple option might be to just only use the first group that is defined with server loads, with no fallback. We probably don't want vslow dump queries falling back to the main DB servers anyway. Multiple servers can already be configured for group if redundancy is desired.
Change 889650 had a related patch set uploaded (by Aaron Schulz; author: Aaron Schulz):
[mediawiki/core@master] rdbms: simplify query group selection in LoadBalancer::getConnection()
Change 889650 merged by jenkins-bot:
[mediawiki/core@master] rdbms: simplify query group selection in LoadBalancer::getConnection()