Details
-
Bug
-
Resolution: Fixed
-
Major
-
4.0.0
-
Security Level: Public
-
Untriaged
-
Release Note
Description
Steps to reproduce:
- set up single node cluster
- set time 2 hours back
- load data in membase
- observe that nothing is getting persisted to disk
- set time 2 hours forward
- observe that is started persisting data
Accoring to Dustin, ep-engine's dispatcher uses pthread_cond_timedwait which uses absolute value of wall-clock time. POSIX allows to configure waiting using monotonic clock, but that feature is optional.
The proposal for now is to avoid large wall-clock changes. Users should run ntpd to keep node's time in sync, but membase service should be started after initial clock synchronization is done.