[
https://issues.apache.org/jira/browse/HBASE-11830?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14534620#comment-14534620
]
Stephen Yuan Jiang commented on HBASE-11830:
--------------------------------------------
The assert in the tests seems invalid to me. It assumed that the test would
run fast between 2 getNextSleepInterval() calls - less than 1 millisecond. It
just put the bet that the test machine is fast enough for test to pass.
The formula of "sleepTicks = shouldTillTo - now" is "sleepTicks =
this.cycleStartTick + (long)(cycles * 100) -
EnvironmentEdgeManager.currentTime()" - For a 100 cycles, the test expects
10000ms or 9999ms sleepTicks return - which bet that this.cycleStartTick that
set earlier almost equal to EnvironmentEdgeManager.currentTime(). In a virtual
box or a slow machine, this could be guaranteed.
[~fenghh] and [~jmspaggi], I think we should relax the assertion - if we go
extreme, anything greater than 0 sleepTicks should be fine (eg.
"assertTrue(ticks2 > 0)"); we can also go some middle ground (eg.
"assertTrue(ticks2 > 90% * 10000)").
The following is more details of my analysis:
- the test (I added some comments to highlight the portion that I am talking
about):
{code}
// throttle bandwidth is 100 and 10 bytes/cycle respectively
ReplicationThrottler throttler1 = new ReplicationThrottler(100);
ReplicationThrottler throttler2 = new ReplicationThrottler(10);
long ticks1 = throttler1.getNextSleepInterval(1000);
=> long ticks2 = throttler2.getNextSleepInterval(1000); //<-First Call-> set
start time//
// 1. the first push size is 1000, though 1000 bytes exceeds 100/10
// bandwidthes, but no sleep since it's the first push of current
// cycle, amortizing occurs when next push arrives
assertEquals(0, ticks1);
assertEquals(0, ticks2);
throttler1.addPushSize(1000);
throttler2.addPushSize(1000);
ticks1 = throttler1.getNextSleepInterval(5);
=> ticks2 = throttler2.getNextSleepInterval(5); //<-Second Call-> fetch
sleepTicks based on First Call, expect <1ms between calls //
// 2. when the second push(5) arrives and throttling(5) is called, the
// current cyclePushSize is 1000 bytes, this should make throttler1
// sleep 1000/100 = 10 cycles = 1s and make throttler2 sleep 1000/10
// = 100 cycles = 10s before the second push occurs -- amortize case
// after amortizing, both cycleStartTick and cyclePushSize are reset
assertTrue(ticks1 == 1000 || ticks1 == 999);
assertTrue(ticks2 == 10000 || ticks2 == 9999); //<-ASSERT FIRE->//
{code}
- the code (I removed some code and comments; and I highlight the portion that
I want to emphasize):
{code}
public long getNextSleepInterval(final int size) {
...
long sleepTicks = 0;
long now = EnvironmentEdgeManager.currentTime();
if ((double)this.cyclePushSize > bandwidth) {
double cycles = Math.ceil((double)this.cyclePushSize / bandwidth);
long shouldTillTo = this.cycleStartTick + (long)(cycles * 100);
if (shouldTillTo > now) {
=> sleepTicks = shouldTillTo - now; // <-The test expects that the Second
Call land here->//
} else {
this.cycleStartTick = now;
}
this.cyclePushSize = 0;
} else {
long nextCycleTick = this.cycleStartTick + 100; //a cycle is 100ms
if (now >= nextCycleTick) {
=> this.cycleStartTick = now; //<--Set by First Call ->//
this.cyclePushSize = 0;
} else if (this.cyclePushSize > 0 &&
...
}
}
return sleepTicks;
}
{code}
> TestReplicationThrottler.testThrottling failed on virtual boxes
> ---------------------------------------------------------------
>
> Key: HBASE-11830
> URL: https://issues.apache.org/jira/browse/HBASE-11830
> Project: HBase
> Issue Type: Bug
> Components: test
> Environment: kvm with Centos 6.5, openjdk1.7
> Reporter: Sergey Soldatov
> Assignee: Stephen Yuan Jiang
> Priority: Minor
>
> during test runs TestReplicationThrottler.testThrottling sometimes fails with
> assertion
> testThrottling(org.apache.hadoop.hbase.replication.regionserver.TestReplicationThrottler)
> Time elapsed: 0.229 sec <<< FAILURE!
> java.lang.AssertionError: null
> at org.junit.Assert.fail(Assert.java:86)
> at org.junit.Assert.assertTrue(Assert.java:41)
> at org.junit.Assert.assertTrue(Assert.java:52)
> at
> org.apache.hadoop.hbase.replication.regionserver.TestReplicationThrottler.testThrottling(TestReplicationThrottler.java:69)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
