hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Seraph Imalia <ser...@eisp.co.za>
Subject Re: RAM Problems - Keeps Crashing
Date Wed, 04 Jan 2012 12:45:42 GMT

On 03 Jan 2012, at 6:39 PM, Stack wrote:

> On Wed, Dec 28, 2011 at 6:27 AM, Seraph Imalia <seraph@eisp.co.za> wrote:
>> After updating from 0.20.6 to 0.90.4, we have been having serious RAM issues.  I
had hbase-env.sh set to use 3 Gigs of RAM with 0.20.6 but with 0.90.4 even 4.5 Gigs seems
not enough.  It does not matter how much load the hbase services are under, it just crashes
after 24-48 hours.
> What kind of a 'crash' is it?  Is it OOME, or JVM seg faulting or just
> a full GC making the RS look like its gone away?

The crash seems slightly different each time (which I suppose is consistent with running out
of RAM).  When our monitoring system alerts me to the problem and I log into the 3 servers,
sometimes the regionservers on dynobuntu10 and dynobuntu12 have already shutdown and the last
thing in their logs says that the Shutdown Hook finished.  The regionsever on dynobuntu17
(which also has the master running) is usually frozen with the last item in the log being
10-20 minutes prior.  

I then run bin/stop-hbase.sh on the dynobuntu17: if the regionservers on dynobuntu10 or dynobuntu12
are still running, sometimes they shutdown gracefully whilst other times the logs just show
Shutdown Hook Initiated and then nothing more.  The master then keeps logging which servers
it is waiting on to shutdown.  I leave it like that for about 5-10 minutes allowing any processes
that are still alive to do as much as they can before I do a kill -9.

That said, for the latest crash: when I logged in, the regionservers on dynobuntu10 and dynobuntu12
had shutdown already, and when I ran bin/stop-hbase.sh on the master, everything shutdown
gracefully (kill -9 was not necessary) - this is the first time this has happened so effortlessly.

>>  The only difference the load makes is how quickly the services crash.  Even over
this holiday season with our lowest load of the year, it crashes just after 36 hours of being
started.  To fix it, I have to run the stop-hbase.sh command, wait a while and kill -9 any
hbase processes that have stopped outputting logs or stopped responding, and then run start-hbase.sh
> The process is deadlocked?   IIRC, 0.90.4 had a possible deadlock.
> You could try 0.90.5.

Sometimes yes, my answer above gives more detail.  
Nice - didn't notice 0.90.5 had been released, I will try that next!

> I took a look at some of the logs.  They do not run from server start
> because i do not see the ulimit output in there.  I'd like to see
> that.

Sorry, I see that now :(.  I have put the logs for the last two crashes here (it's 2.5 Megs):
One crash was around 19:30 yesterday and the second was at 12:50 today.

> Looking at dynobuntu10, I see some interesting 'keys':
> 2011-12-28 15:25:53,297 INFO
> org.apache.hadoop.hbase.regionserver.HRegionServer: Received request
> to open region:
> UrlIndex,http://www.hellopeter.com/write_report_preview.php?inclination=1&company=kalahari.net&countryid=168&location=cape
> town&industryid=14&person=&problem=out of stock&other=&headline=why

> advertise goods online and you cannot deliver%29&incident=i purchased
> goods online that were supposedly in stock on the 5th october. 2010.
> after numerous phone calls i was promised that i would receive the
> ordered goods on the 20th october 2010. this has not happened to date.
> i spoke with them today and they promised to answer my queries on
> 21st october2010. how can you run a online busines ans sell %22we dont
> have stock%22%3a this is the easy way out as we have no proof of
> that%0d%0ait is just common curtousy to return a phone call. they have
> had my money in their bank account for 15 days. this seems like a
> ****. they could be reaping interest on thousands of peoples money.
> easy way of making money.%0d%0akalahari. net are in a comfort zone.
> they need to realize that customers are king%0d%0athey reimburse my
> money. i paid bank charges and transfer fees. what about this. my
> unnessessary phone calls. do they reinburse this.%0d%0acome on stop
> taking the innocent public for a ride with your sweet
> talk.&incidentcharsleft=270&incident_day_select=21&incident_month_select=10&incident_year_select=2010&incident_hour_select=11&incident_min_select=45&incident_ampm_select=pm&policyno=3573210
> %2f3573310 &cellno=%2b27
> 766881896&preview=preview,1308921597915.1827414390
> Thats a single key.  It looks like you have an issue in your crawler's
> url extraction facility.

Yeah, that URL actually exists, but I can see how that can be a problem to use as a key. 
Not sure what to do here, perhaps we should exclude URL's like this - or perhaps your hashing
idea below will solve that.  I don't really know enough about hashing to make the call though
- is it not possible to run into duplicate keys using e.g. an MD5 Hash? - The MD5 Hash of
the above URL is: 8f157d290ceeacedb6c1be133f1ca153 - it seems logical to me that a string
that small cannot possibly be unique given that the URL was originally 1431 characters long.
 What is your opinion on this? - I will be doing some more research on this though.  Perhaps
there is a Hash-type that is more unique that you could suggest for our purposes (but keeping
in mind our ad delivery servers will need to hash the URL before querying hBase so it needs
to be fast and not resource intensive)?

> If you have lots of URLs like the above, my guess is that you have
> massive indices.  Look at a regionserver and see how much RAM the
> indexes take up?

Yeah, looks pretty high, it is currently half the MaxHeap on a fresh start... 

Below is what it is now, I have just disabled the block cache after the last crash as you
suggested to try keep it stable until we have a real fix.  With the Block cache at the default
of 25% (1 Gig) and the IndexSize at around 2 Gigs, that only leaves 1 Gig for everything else
:( which is not much.

dynobuntu10: requests=93, regions=225, stores=225, storefiles=354, storefileIndexSize=2239,
memstoreSize=35, compactionQueueSize=0, flushQueueSize=0, usedHeap=2639, maxHeap=4087, blockCacheSize=0,
blockCacheFree=0, blockCacheCount=0, blockCacheHitCount=0, blockCacheMissCount=0, blockCacheEvictedCount=0,
blockCacheHitRatio=0, blockCacheHitCachingRatio=0

dynobuntu12: requests=305, regions=225, stores=225, storefiles=435, storefileIndexSize=2004,
memstoreSize=31, compactionQueueSize=0, flushQueueSize=0, usedHeap=2321, maxHeap=4087, blockCacheSize=0,
blockCacheFree=0, blockCacheCount=0, blockCacheHitCount=0, blockCacheMissCount=0, blockCacheEvictedCount=0,
blockCacheHitRatio=0, blockCacheHitCachingRatio=0

dynobuntu17: requests=51, regions=226, stores=226, storefiles=410, storefileIndexSize=2046,
memstoreSize=36, compactionQueueSize=0, flushQueueSize=0, usedHeap=2927, maxHeap=4087, blockCacheSize=0,
blockCacheFree=0, blockCacheCount=0, blockCacheHitCount=0, blockCacheMissCount=0, blockCacheEvictedCount=0,
blockCacheHitRatio=0, blockCacheHitCachingRatio=0

> In dynoubuntu12 I see an OOME.  Interestingly, the OOME is while
> trying to read in a file's index on:
> 2011-12-28 15:26:50,310 DEBUG
> org.apache.hadoop.hbase.regionserver.HRegion: Opening region: REGION
> => {NAME => 'UrlIndex,http://media.imgbnr.com/images/prep_ct.php?imgfile=4327_567146_7571713_250_300.html&partnerid=113471&appid=35229&subid=&advertiserid=567146&keywordid=42825417&type=11&uuid=e11ac4bea82d42838fde8eb306fbc354&keyword=www.&matchedby=c&ct=cpi&wid=5008233&size=300x250&lid=7571713&cid=230614&cc=us&rc=in&mc=602&dc=0&vt=1275659190365&refurl=mangafox.com&clickdomain=,1283006905877',
> STARTKEY => 'http://media.imgbnr.com/images/prep_ct.php?imgfile=4327_567146_7571713_250_300.html&partnerid=113471&appid=35229&subid=&advertiserid=567146&keywordid=42825417&type=11&uuid=e11ac4bea82d42838fde8eb306fbc354&keyword=www.&matchedby=c&ct=cpi&wid=5008233&size=300x250&lid=7571713&cid=230614&cc=us&rc=in&mc=602&dc=0&vt=1275659190365&refurl=mangafox.com&clickdomain=',
> ENDKEY => 'http://media.imgbnr.com/images/prep_ct.php?imgfile=6966_567146_7571715_90_728.html&partnerid=113474&appid=35224&subid=&advertiserid=567146&keywordid=42825616&type=11&uuid=6178294088f545ab938c403be5b7c957&keyword=www.&matchedby=c&ct=cpi&wid=5008236&size=728x90&lid=7571715&cid=230615&cc=us&rc=ny&mc=501&dc=0&vt=1275772980357&refurl=worldstarhiphop.com&clickdomain=',
> ENCODED => 1246560666, TABLE => {{NAME => 'UrlIndex', INDEXES =>
> 'indexUrlUIDUrlIndex_Family:urluid=org.apache.hadoop.hbase.client.tableindexed.IndexKeyGenerator:com.entelligence.tools.hbase.index.UniqueIndexKeyGeneratorUrlIndex_Family:urluidorg.apache.hadoop.io.Writable0org.apache.hadoop.io.ObjectWritable$NullInstance'org.apache.hadoop.io.WritableComparableindexHostLocationsUrlIndex_Family:hostUrlIndex_Family:locationcodes=org.apache.hadoop.hbase.client.tableindexed.IndexKeyGeneratorCorg.apache.hadoop.hbase.client.tableindexed.SimpleIndexKeyGeneratorUrlIndex_Family:hostorg.apache.hadoop.io.Writable0org.apache.hadoop.io.ObjectWritable$NullInstance'org.apache.hadoop.io.WritableComparableindexLocationCodesUrlIndex_Family:locationcodes=org.apache.hadoop.hbase.client.tableindexed.IndexKeyGeneratorCorg.apache.hadoop.hbase.client.tableindexed.SimpleIndexKeyGeneratorUrlIndex_Family:locationcodesorg.apache.hadoop.io.Writable0org.apache.hadoop.io.ObjectWritable$NullInstance'org.apache.hadoop.io.WritableComparable
> FAMILIES => [{NAME => 'UrlIndex_Family', query => '', datelastspidered
> => '', path => '', BLOOMFILTER => 'NONE', TTL => '2147483647',
> datenextspider => '', daylastaccesed => '', host => '', originalurl =>
> '', locationcodes => '', extension => '', IN_MEMORY => 'false',
> COMPRESSION => 'LZO', VERSIONS => '3', protocol => '', failCount =>
> '', datediscovered => '', contentDiff => '', BLOCKSIZE => '65536',
> datelastmodified => '', urluid => '', BLOCKCACHE => 'true',
> channeluids => ''}]}}
> What is that INDEXES thing in the above schema?  Is that some
> secondary indexing thing you have going on?

Yes, it is a secondary Index we created. Basically, we serve ads and ads are queued per URL.
 When we discover a new URL, we add it to HBase and give it a GUID which is stored as a column.
 Other servers build lists of ads for each URL and store it against the GUID of the URL. 
So when a request comes in for ads, we use HBase to lookup the URL and get the GUID so that
it can then know which ads to show.  This is all handled by the main URLIndex Table.  BUT,
sometimes and far less often, we have a situation where we have the GUID but need to look
up the URL - so we have another table where the GUID is the rowKey and the URL is a column.
 We also do this for hosts and for Channels (this is what we call a place where ads show).

> You might take a look at the files under UrlIndex/1246560666 in the
> UrlIndex column family..... Print out their meta data and see what
> size indices you have.   See
> http://hbase.apache.org/book.html#rowkey.design in the book.  It has
> some pointers and some talk on issues you may be running into.

root@dynobuntu17:/opt/hadoop-0.20.2# bin/hadoop fs -ls /hbase/UrlIndex/1246560666/UrlIndex_Family
Found 1 items
-rw-r--r--   2 root supergroup  199115380 2010-09-10 01:52 /hbase/UrlIndex/1246560666/UrlIndex_Family/6442966743799940481

After reading that, it seems clear we need to make some minor changes to our Table Design,
unfortunately it means creating a new table and copying rows to the new table - not a fun
process because we can't be down whilst doing it, so we'll have to write some good code to
ease the process - doable, but not fun.  I am hoping that upgrading the version to 0.90.5
and disabling the block cache buys us about a month so we have the time to plan it properly.

>> Attached are my logs from the latest "start-to-crash".  There are 3 servers and hbase
is being used for storing URL's - 7 client servers connect to hbase and perform URL Lookups
at about 40 requests per second (this is the low load over this holiday season).  If the URL
does not exist, it gets added.  The Key on the HTable is the URL and there are a few fields
stored against it - e.g. DateDiscovered, Host, Script, QueryString, etc.
> Do you have to scan the URLs in order or by website?  If not, you
> might have a key that is a hash of the URL (and keep actual URL as
> column data).

Yes sometimes we need to do scans like that - but only for a manual investigation, not during
normal operation.  We may be able to get by as long as we can come up with a plan for how
we can find the url's for a particular website.  I am concerned about the uniqueness of a
hash.  I see there are lots of different hashes.  Will there be uniqueness issues? - we can't
have two URL's having the same hash.

>> Each server has a hadoop datanode and an hbase regionserver and 1 of the servers
additionally has the namenode, master and zookeeper.  On first start, each regionserver uses
2 Gigs (usedHeap) and as soon as I restart the clients, the usedHeap slowly climes until it
reaches the maxHeap and shortly after that, the regionservers start crashing - sometimes they
actually shutdown gracefully by themselves.
> Are the URL lookups totally random?  If so, turn off the block cache.
> That'll get you some more memory.

Yes, pretty random and as we grow, it will get more random.  I have disabled the block cache
- and looking at the heap stats which I pasted above, it seems like it will buy us some time
to make some long-term changes - I will keep you updated here.

> Add more servers too to spread the load if you can afford it.  Things
> tend to run smoother once you get above 5 servers or so.

We currently have 4 instances of HBase - 2 each have 5 servers and are used for Ad Delivery
Log Storage and the other two are used for URL lookups and each have 3 servers.  I will struggle
to get our finance guys to approve more servers for hBase, but if that is my only option I
will definitely try :)

Coincidently, the 2 instances used for Ad Delivery log storage are down at the moment, but
it is because we are having stability issues with Ubuntu Server - they periodically do a memory
dump and shut down - even if nothing is running on them.  I have to tackle that problem too
pretty soon.  In the meantime, MySQL is taking up the slack, but we will quickly run into
performance issues if we don't fix that soon.  But anyway, at the moment I don't need your
help with those servers because it does not seem to be hBase or hadoop causing the crashes.
 I am tackling this URL Server problem first.

>> Originally, we had hbase.regionserver.handler.count set to 100 and I have now removed
that to leave it as default which has not helped.
>> We have not made any changes to the clients and we have a mirrored instance of this
in our UK Data Centre which is still running 0.20.6 and servicing 10 clients currently at
over 300 requests per second (again low load over the holidays) and it is 100% stable.
>> What do I do now? - your website says I cannot downgrade?
> That is right.
> Lets get this stable again.
> St.Ack

Thanks for your help so far.  I have already disabled the block cache (which I am sure will
show an immediate improvement) and I will schedule an upgrade of hBase to 0.90.5 during this
week and then monitor it.  If you do have some knowledge about the uniqueness of an MD5 Hash,
please share it with me if you have the time? - it will help me whilst I plan the changes
we need to make to the table structure.


View raw message