hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From lars hofhansl <lhofha...@yahoo.com>
Subject Re: Extremely long flush times
Date Sat, 11 Aug 2012 06:43:13 GMT
Ran your test code (thanks Carlos).

Found two things:
1. Store.internalFlushCache(...) should be calling StoreScanner.next(List<KeyValue>,
int limit) - currently it does not set a limit.(But this is not the problem).

2. With jstack I found that the code is stuck in a loop in Memstore.MemstoreScanner.getNext(...)

Here's the relevant part of the jstack:
"IPC Server handler 6 on 60020" daemon prio=10 tid=0x00007f0574625000 nid=0x720c runnable
   java.lang.Thread.State: RUNNABLE
        at org.apache.hadoop.hbase.regionserver.MemStore$MemStoreScanner.getNext(MemStore.java:726)
        at org.apache.hadoop.hbase.regionserver.MemStore$MemStoreScanner.seekInSubLists(MemStore.java:761)
        - locked <0x00000000c4a8a860> (a org.apache.hadoop.hbase.regionserver.MemStore$MemStoreScanner)
        at org.apache.hadoop.hbase.regionserver.MemStore$MemStoreScanner.reseek(MemStore.java:800)
        - locked <0x00000000c4a8a860> (a org.apache.hadoop.hbase.regionserver.MemStore$MemStoreScanner)
        at org.apache.hadoop.hbase.regionserver.NonLazyKeyValueScanner.doRealSeek(NonLazyKeyValueScanner.java:54)
        at org.apache.hadoop.hbase.regionserver.KeyValueHeap.generalizedSeek(KeyValueHeap.java:299)
        at org.apache.hadoop.hbase.regionserver.KeyValueHeap.reseek(KeyValueHeap.java:244)
        at org.apache.hadoop.hbase.regionserver.StoreScanner.reseek(StoreScanner.java:522)
        - eliminated <0x00000000ccb54860> (a org.apache.hadoop.hbase.regionserver.StoreScanner)
        at org.apache.hadoop.hbase.regionserver.StoreScanner.next(StoreScanner.java:403)
        - locked <0x00000000ccb54860> (a org.apache.hadoop.hbase.regionserver.StoreScanner)
        at org.apache.hadoop.hbase.regionserver.KeyValueHeap.next(KeyValueHeap.java:127)
        at org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl.nextInternal(HRegion.java:3459)
        at org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl.next(HRegion.java:3406)
        - locked <0x00000000c59ee610> (a org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl)
        at org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl.next(HRegion.java:3423)
        - locked <0x00000000c59ee610> (a org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl)
        at org.apache.hadoop.hbase.regionserver.HRegion.get(HRegion.java:4171)
        at org.apache.hadoop.hbase.regionserver.HRegion.get(HRegion.java:4144)
        at org.apache.hadoop.hbase.regionserver.HRegionServer.get(HRegionServer.java:1958)
        at sun.reflect.GeneratedMethodAccessor12.invoke(Unknown Source)
        at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
        at java.lang.reflect.Method.invoke(Method.java:601)
        at org.apache.hadoop.hbase.ipc.WritableRpcEngine$Server.call(WritableRpcEngine.java:364)
        at org.apache.hadoop.hbase.ipc.HBaseServer$Handler.run(HBaseServer.java:1389)

At the same time I find that flush cannot finish:

"regionserver60020.cacheFlusher" daemon prio=10 tid=0x00007f05749ab000 nid=0x71fe waiting
for monitor entry [0x00007f05677f6000]
   java.lang.Thread.State: BLOCKED (on object monitor)
        at org.apache.hadoop.hbase.regionserver.StoreScanner.updateReaders(StoreScanner.java:443)
        - waiting to lock <0x00000000ccb54860> (a org.apache.hadoop.hbase.regionserver.StoreScanner)
        at org.apache.hadoop.hbase.regionserver.Store.notifyChangedReadersObservers(Store.java:904)
        at org.apache.hadoop.hbase.regionserver.Store.updateStorefiles(Store.java:893)
        at org.apache.hadoop.hbase.regionserver.Store.access$600(Store.java:107)
        at org.apache.hadoop.hbase.regionserver.Store$StoreFlusherImpl.commit(Store.java:2291)
        at org.apache.hadoop.hbase.regionserver.HRegion.internalFlushcache(HRegion.java:1455)
        at org.apache.hadoop.hbase.regionserver.HRegion.internalFlushcache(HRegion.java:1353)
        at org.apache.hadoop.hbase.regionserver.HRegion.flushcache(HRegion.java:1294)
        at org.apache.hadoop.hbase.regionserver.MemStoreFlusher.flushRegion(MemStoreFlusher.java:406)
        at org.apache.hadoop.hbase.regionserver.MemStoreFlusher.flushRegion(MemStoreFlusher.java:380)
        at org.apache.hadoop.hbase.regionserver.MemStoreFlusher.run(MemStoreFlusher.java:243)
        at java.lang.Thread.run(Thread.java:722)

Both StoreScanner.updateReaders and StoreScanner.reseek are synchronized.

So the problem seems to be that MemStoreScanner loops forever in getNext(...). I took a jstack
a bunch of times during execution, this always shows up.
Need to dig a bit more, I do not see a good way to deal with this, yet.

-- Lars

----- Original Message -----
From: "Espinoza,Carlos" <espinozca@oclc.org>
To: user@hbase.apache.org
Sent: Friday, August 10, 2012 3:12 PM
Subject: RE: Extremely long flush times

We were able to replicate this behavior in a pseudo-distributed hbase
(hbase-0.94.1) environment. We wrote a test program that creates a test
table "MyTestTable" and populates it with random rows, then it creates a
row with 60,000 columns and repeatedly updates it. Each column has a 18
byte qualifier and a 50 byte value. In our tests, when we ran the
program, we usually never got beyond 15 updates before it would flush
for a really long time. The rows that are being updated are about 4MB
each (minues any hbase metadata).

It doesn't seem like it's caused by GC. I turned on gc logging, and
didn't see any long pauses. This is the gc log during the flush.

This is the regionserver log with debug on during the same flush

This is the test program we wrote. 

You should be able to just compile it, and run it against a running
HBase cluster.
$ java TestTable


-----Original Message-----
From: jdcryans@gmail.com [mailto:jdcryans@gmail.com] On Behalf Of
Jean-Daniel Cryans
Sent: Monday, August 06, 2012 8:04 PM
To: user@hbase.apache.org
Subject: Re: Extremely long flush times

Hi Ron,

I think this is caused by GC, for example there's 7 minutes between the
last blocking message and the end of the flush:

2012-08-06 11:44:36,348 INFO
org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for 'IPC
Server handler 42 on 9009' on region
memstore size 271.9m is >= than blocking 256.0m size
2012-08-06 11:51:28,139 INFO
org.apache.hadoop.hbase.regionserver.HRegion: Finished memstore flush of
~155.4m/162905016, currentsize=116.5m/122172880 for region
Works,73658752:eng,1337915462169.1cfc92481a74ec278337871563d189d6. in
501337ms, sequenceid=1792264092, compaction requested=true

During that time I would expect that clients got socket timeouts, tried
to write again, and data would just pile up in the RPC queue.

Do you have the gc log handy? Also did you set the zookeeper session
timeout higher than the default of 3 minutes?



On Mon, Aug 6, 2012 at 2:37 PM, Buckley,Ron <buckleyr@oclc.org> wrote:
> We ran into this situation today on our production instance, I'm 
> wondering if anyone has seen or knows a way around it.
> We were repeatedly updating the same 60,000 column wide row.  With 
> each update, we were updating nearly every column of the row. Each 
> update generated nearly 9 MB in the WAL.
> This worked OK most of the time. We saw frequent messages like this in

> the region server log, indicating a 2.6 second flush time:
> 2012-08-06 11:42:52,316 INFO
> org.apache.hadoop.hbase.regionserver.HRegion: Finished memstore flush 
> of ~155.4m/162931400, currentsize=2.7k/2760 for region 
> Works,73658752:eng,1337915462169.1cfc92481a74ec278337871563d189d6. in 
> 2610ms, sequenceid=1792263953, compaction requested=false
> However, sometimes we would see a flush times around 8 - 10 Minutes.
> 2012-08-06 11:51:28,139 INFO
> org.apache.hadoop.hbase.regionserver.HRegion: Finished memstore flush 
> of ~155.4m/162905016, currentsize=116.5m/122172880 for region 
> Works,73658752:eng,1337915462169.1cfc92481a74ec278337871563d189d6. in 
> 501337ms, sequenceid=1792264092, compaction requested=true
> During these long flushes, we would end up taking all the ipc handler 
> slots into that region server that we had configured (50).  It looked 
> to our application like the region server was unavailable.
> After the long delay, the region server would continue running fine, 
> but then do it again a few minutes later.
> I thought it was most likely something with Garbage Collection, so I 
> got a stack trace during one of the events.  The trace doesn't look 
> like GC was running.
> http://pastebin.com/ANBiUzX4  - Stack Trace when the region server was

> unreachable
> http://pastebin.com/naTYQxZR  - Region Server Log during the same 
> interval
> Versions:
> HBase 0.92.1, r1298924
> Hadoop 0.20.2-cdh3u3
> ----------------------------------------------------------------------
> Ron Buckley
> x6365

View raw message