Redundant power supplies are probably your next-best bet for running without fsync (hsync). Something that can prevent a node from going down hard will mitigate this issue for the most part.

The importance of this is often a multi-variable equation. The small chance for data loss that exists with hflush can be a non-issue for some (e.g. ability to replay original data, acceptability of minor data loss, knowledge to rebuild hbase:meta as required)

On 12/7/17 6:37 PM, Daniel Połaczański wrote:
Hi,
fsync mode on WAL is not supported currenlty. So theoritically it is
possible that in case of power failure the data is lost because DataNodes
didn't flush it to physical disk.

I know that probability is small and data center should have additional
power source, but still it is possible. How do you deal with it? Do you
care about it? How to check after the disaster later from which moment the
data is lost? How often DataNodes flush buffers to disk?
What is your solution?

Regards

Reply via email to