[KUDU-503] Regression: always sync-on-close for data blocks makes tablet creation too slow

• Type: Sub-task
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: M4.5
• Component/s: fs
• Labels: data-scalability
• Assignee: Unassigned
• Reporter: Jean-Daniel Cryans
• Created: 12/Sep/14 16:09
• Updated: 06/Mar/17 21:20

Description

When this change (http://gerrit.sjc.cloudera.com:8080/#/c/4231/) was pushed a bunch of stuff got slower, for example: http://gerrit.sjc.cloudera.com:8080/#/c/4253/.

Tablet creation also got a lot slower, we used to be able to create 700 tablets with num_replicas=3 but now we're always missing ~50 (because the master drops the calls on the floor, different issue).

enable_data_block_fsync can be disabled to get the performance back.

On a single node, tablet creation with the flag turned off takes ~100ms whereas with the default value it takes at least 700ms (saw it taking seconds).