Introduce startup-silent-period mechanism to avoid partial assignments

[zuston]

What changes were proposed in this pull request?

Introduce startup-silent-period mechanism to avoid partial assignments

Why are the changes needed?

When changing some coordinator's conf and then restart, coordinator will accept client getAssignment request immediately, but it will serve for jobs request based on the partial registered shuffle-servers, which will make some jobs gotten not enough required shuffle-servers and then slow the running speed.

I think we should make coordinator wait for more than one shuffle-server heartbeat interval before serving for client. During out-of-service, requests from client will fallback to slave coordinator.

Does this PR introduce any user-facing change?

Yes

How was this patch tested?

UTs

[zuston]

PTAL @jerqi

[codecov-commenter]

Codecov Report

Merging #247 (63af987) into master (c89f95c) will increase coverage by 0.02%. The diff coverage is 66.66%.

@@             Coverage Diff              @@
##             master     #247      +/-   ##
============================================
+ Coverage     59.16%   59.18%   +0.02%     
- Complexity     1340     1343       +3     
============================================
  Files           163      163              
  Lines          8810     8837      +27     
  Branches        833      835       +2     
============================================
+ Hits           5212     5230      +18     
- Misses         3332     3339       +7     
- Partials        266      268       +2     

Impacted Files	Coverage Δ
...he/uniffle/coordinator/CoordinatorGrpcService.java	2.31% <0.00%> (-0.03%)	:arrow_down:
...ache/uniffle/coordinator/SimpleClusterManager.java	86.61% <53.33%> (-4.46%)	:arrow_down:
...rg/apache/uniffle/coordinator/CoordinatorConf.java	97.26% <100.00%> (+0.20%)	:arrow_up:

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more

[jerqi]

Is startup-silent-period a good name? What's name of similar mechanism of Yarn?

[zuston]

Is startup-silent-period a good name? What's name of similar mechanism of Yarn?

No such config in Yarn. Safe mode is for HDFS. I think it's not proper in Uniffle scenarios

Do u have any ideas?

[jerqi]

Is startup-silent-period a good name? What's name of similar mechanism of Yarn?

No such config in Yarn. Safe mode is for HDFS. I think it's not proper in Uniffle scenarios

Do u have any ideas?

It's ok for Uniffle. I don't have another good ideas.

[zuston]

It's ok for Uniffle. I don't have another good ideas.

Got it. I will rename to safemode.

[jerqi]

It's ok for Uniffle. I don't have another good ideas.

Got it. I will rename to safemode.

I means that startup-silent-period is ok.

[zuston]

OK. Do u have any other advice? @jerqi

[zuston]

Updated.

1. Add the doc of these configs
2. Explain why make the startup-slient-period stop default.

[zuston]

Error:  Tests run: 6, Failures: 1, Errors: 0, Skipped: 0, Time elapsed: 17.998 s <<< FAILURE! - in org.apache.uniffle.test.CoordinatorGrpcTest
Error:  rpcMetricsTest  Time elapsed: 1.035 s  <<< FAILURE!
org.opentest4j.AssertionFailedError: expected: <0.0> but was: <1.0>

CI failed. https://github.com/apache/incubator-uniffle/actions/runs/3216468616/jobs/5258387024 #244

Rerun it.

[jerqi]

Unrelated flaky test failed, I will rerun the flaky test