Lineland

One of the more ambiguous things in Hadoop is block replication: it happens automatically and you should not have to worry about it. HBase relies on it 100% to provide the data safety as it stores its files into the distributed file system. While that works completely transparent, one of the more advanced questions asked though is how does this affect performance? This usually arises when the user starts writing MapReduce jobs against either HBase or Hadoop directly. Especially with larger data being stored in HBase, how does the system take care of placing the data close to where it is needed? This is referred to data locality and in case of HBase using the Hadoop file system (HDFS) there may be doubts how that is working.

First let's see how Hadoop handles this. The MapReduce documentation advertises the fact that tasks run close to the data they process. This is achieved by breaking up large files in HDFS into smaller chunks, or so called blocks. That is also the reason why the block size in Hadoop is much larger than you may know them from operating systems and their file systems. Default setting is 64MB, but usually 128MB is chosen, if not even larger when you are sure all your files are larger than a single block in size. Each block maps to a task run to process the contained data. That also means larger block sizes equal fewer map tasks to run as the number of mappers is driven by the number of blocks that need processing. Hadoop knows where blocks are located and runs the map tasks directly on the node that hosts it (actually one of them as replication means it has a few hosts to chose from). This is how it guarantees data locality during MapReduce.

Back to HBase. When you have arrived at that point with Hadoop and you now understand that it can process data locally you start to question how this may work with HBase. If you have read my post on HBase's storage architecture you saw that HBase simply stores files in HDFS. It does so for the actual data files (HFile) as well as its log (WAL). And if you look into the code it simply uses FileSystem.create(Path path) to create these. When you then consider two access patterns, a) direct random access and b) MapReduce scanning of tables, you wonder if care was taken that the HDFS blocks are close to where they are read by HBase.

One thing upfront, if you do not co-share your cluster with Hadoop and HBase but instead employ a separate Hadoop as well as a stand-alone HBase cluster then there is no data locality - and it can't be. That equals to running a separate MapReduce cluster where it would not be able to execute tasks directly on the datanode. It is imperative for data locality to have them running on the same cluster, Hadoop (as in the HDFS), MapReduce and HBase. End of story.

OK, you them all co-located on a single (hopefully larger) cluster? Then read on. How does Hadoop figure out where data is located as HBase accesses it. Remember the access pattern above, both go through a single piece of software called a RegionServer. Case a) uses random access patterns while b) scans all contiguous rows of a table but does so through the same API. As explained in my referenced post and mentioned above, HBase simply stores files and those get distributed as replicated blocks across all data nodes of the HDFS. Now imagine you stop HBase after saving a lot of data and restarting it subsequently. The region servers are restarted and assign a seemingly random number of regions. At this very point there is no data locality guaranteed - how could it be?

The most important factor is that HBase is not restarted frequently and that it performs house keeping on a regular basis. These so called compactions rewrite files as new data is added over time. All files in HDFS once written are immutable (for all sorts of reasons). Because of that, data is written into new files and as their number grows HBase compacts them into another set of new, consolidated files. And here is the kicker: HDFS is smart enough to put the data where it is needed! How does that work you ask? We need to take a deep dive into Hadoop's source code and see how the above FileSystem.create(Path path) that HBase uses works. We are running on HDFS here, so we are actually using DistributedFileSystem.create(Path path) which looks like this:

public FSDataOutputStream create(Path f) throws IOException {
  return create(f, true);
}

It returns a FSDataOutputStream and that is create like so:

public FSDataOutputStream create(Path f, FsPermission permission, boolean overwrite, int bufferSize, short replication, long blockSize, Progressable progress) throws IOException {
  return new FSDataOutputStream(dfs.create(getPathName(f), permission, overwrite, replication, blockSize, progress, bufferSize), statistics);
}

It uses a DFSClient instance that is the "umbilical" cord connecting the client with the NameNode:

this.dfs = new DFSClient(namenode, conf, statistics);

What is returned though is a DFSClient.DFSOutputStream instance. As you write data into the stream the DFSClient aggregates it into "packages" which are then written as blocks to the data nodes. This happens in DFSClient.DFSOutputStream.DataStreamer (please hang in there, we are close!) which runs as a daemon thread in the background. The magic unfolds now in a few hops on the stack, first in the daemon run() it gets the list of nodes to store the data on:

nodes = nextBlockOutputStream(src);

This in turn calls:

long startTime = System.currentTimeMillis();
lb = locateFollowingBlock(startTime);
block = lb.getBlock();
nodes = lb.getLocations();

We follow further down and see that locateFollowingBlocks() calls:

return namenode.addBlock(src, clientName);

Here is where it all comes together. The name node is called to add a new block and the src parameter indicates for what file, while clientName is the name of the DFSClient instance. I skip one more small method in between and show you the next bigger step involved:

public LocatedBlock getAdditionalBlock(String src, String clientName) throws IOException {
  ...
  INodeFileUnderConstruction pendingFile  = checkLease(src, clientName);
  ...
  fileLength = pendingFile.computeContentSummary().getLength();
  blockSize = pendingFile.getPreferredBlockSize();
  clientNode = pendingFile.getClientNode();
  replication = (int)pendingFile.getReplication();

  // choose targets for the new block tobe allocated.
  DatanodeDescriptor targets[] = replicator.chooseTarget(replication, clientNode, null, blockSize);
  ...
}

We are finally getting to the core of this code in the replicator.chooseTarget() call:

private DatanodeDescriptor chooseTarget(int numOfReplicas, DatanodeDescriptor writer, List<Node> excludedNodes, long blocksize, int maxNodesPerRack, List<DatanodeDescriptor> results) {
  
  if (numOfReplicas == 0 || clusterMap.getNumOfLeaves()==0) {
    return writer;
  }
  
  int numOfResults = results.size();
  boolean newBlock = (numOfResults==0);
  if (writer == null && !newBlock) {
    writer = (DatanodeDescriptor)results.get(0); 
  }
  
  try {
    switch(numOfResults) {
    case 0:
      writer = chooseLocalNode(writer, excludedNodes, blocksize, maxNodesPerRack, results);
      if (--numOfReplicas == 0) {
        break;
      }
    case 1:
      chooseRemoteRack(1, results.get(0), excludedNodes, blocksize, maxNodesPerRack, results);
      if (--numOfReplicas == 0) {
        break;
      }
    case 2:
      if (clusterMap.isOnSameRack(results.get(0), results.get(1))) {
        chooseRemoteRack(1, results.get(0), excludedNodes, blocksize, maxNodesPerRack, results);
      } else if (newBlock) {
        chooseLocalRack(results.get(1), excludedNodes, blocksize, maxNodesPerRack, results);
      } else {
        chooseLocalRack(writer, excludedNodes, blocksize, maxNodesPerRack, results);
      }
      if (--numOfReplicas == 0) {
        break;
      }
    default:
      chooseRandom(numOfReplicas, NodeBase.ROOT, excludedNodes, blocksize, maxNodesPerRack, results);
    }
  } catch (NotEnoughReplicasException e) {
    FSNamesystem.LOG.warn("Not able to place enough replicas, still in need of " + numOfReplicas);
  }
  return writer;
}

Recall that we have started with the DFSClient and created a file which was subsequently filled with data. As the blocks need writing out the above code checks first if that can be done on the same host that the client is on, i.e. the "writer". That is "case 0". In "case 1" the code tries to find a remote rack to have a distant replication of the block. Lastly is fills the list of required replicas with local or machines of another rack.

So this means for HBase that as the region server stays up for long enough (which is the default) that after a major compaction on all tables - which can be invoked manually or is triggered by a configuration setting - it has the files local on the same host. The data node that shares the same physical host has a copy of all data the region server requires. If you are running a scan or get or any other use-case you can be sure to get the best performance.

Finally a good overview over the HDFS design and data replication can be found here. Also note that the HBase team is working on redesigning how the Master is assigning the regions to servers. The plan is to improve it so that regions are deployed on the server where most blocks are. This will particularly be useful after a restart because it would guarantee a better data locality right off the bat. Stay tuned!

A quick post explaining how a minimal Katta Lucene Client is set up. I found this was sort of missing from the Katta site and documentation and since I ran into an issue along the way I thought I post my notes here for others who may attempt the same.

First was the question, which of the libs needed to be supplied for a client to use a remote Katta cluster. Please note that I am referring here to a "canonical" setup with a distributed Lucene index (which I created on Hadoop from data in HBase using a MapReduce job). I found these libs needed to be added, the rest is for the server:

katta-core-0.6.rc1.jar
lucene-core-3.0.0.jar
zookeeper-3.2.2.jar
zkclient-0.1-dev.jar
hadoop-core-0.20.1.jar
log4j-1.2.15.jar
commons-logging-1.0.4.jar

Here is the code for the client, please note that this is a simple test app that expects to get the name of the index, the default Lucene search field and query on the command line. I did not add usage info as this is just a proof of concept.

package com.worldlingo.test;

import net.sf.katta.lib.lucene.Hit;
import net.sf.katta.lib.lucene.Hits;
import net.sf.katta.lib.lucene.LuceneClient;
import net.sf.katta.util.ZkConfiguration;
import org.apache.hadoop.io.MapWritable;
import org.apache.hadoop.io.Writable;
import org.apache.lucene.analysis.Analyzer;
import org.apache.lucene.analysis.standard.StandardAnalyzer;
import org.apache.lucene.queryParser.QueryParser;
import org.apache.lucene.search.Query;
import org.apache.lucene.util.Version;

import java.util.Arrays;
import java.util.Map;

public class KattaLuceneClient {

  public static void main(String[] args) {
    try {
      Analyzer analyzer = new StandardAnalyzer(Version.LUCENE_CURRENT);
      Query query = new QueryParser(Version.LUCENE_CURRENT, args[1], analyzer).parse(args[2]);

      // assumes "/katta.zk.properties" available on classpath!
      ZkConfiguration conf = new ZkConfiguration();
      LuceneClient luceneClient = new LuceneClient(conf);
      Hits hits = luceneClient.search(query, Arrays.asList(args[0]).toArray(new String[1]), 99);

      int num = 0;
      for (Hit hit : hits.getHits()) {
        MapWritable mw = luceneClient.getDetails(hit);
        for (Map.Entry<Writable, Writable> entry : mw.entrySet()) {
          System.out.println("[" + (num++) + "] key -> " + entry.getKey() + ", value -> " + entry.getValue());
        }
      }
    } catch (Exception e) {
      e.printStackTrace();
    }
  }

}

The first part is standard Lucene code were we parse the query string with an analyzer. The seconds part is Katta related as it creates a configuration object, which assumes we have a ZooKeeper configuration in the class path. That config only needs to have these lines set:

zookeeper.embedded=false
zookeeper.servers=server-1:2181,server-2:2181

The first line is really only used on the server, so it can be left out on the client. I simply copied the server katta.zk.properties to match my setup. The important line is the second one, which tells the client where the ZooKeeper responsible for managing the Katta cluster is running. With this info the client is able to distribute the search calls to the correct Katta slaves.

Further along we create a LuceneClient instance and start the actual search. Here I simply used no sorting and set the maximum number of hits returned to 99. These two values could be optionally added to the command line parameters but are trivial and not required here - this is a minimal test client after all ;)

The last part of the app is simply printing out the fields and their values of each found document. Please note that Katta is using the low-level Writable class as part of its response. This is not "too" intuitive for the uninitiated. These are actually Text instances so they can safely be convert to text using ".toString()".

Finally, I also checked the test project into my GitHub account for your perusal. Have fun!

I am pleased to invite you to our third Munich Open Hadoop User Group Meeting!

Like always we are looking forward to see everyone again and are welcoming new attendees to join our group. We are enthusiast about all things related to scalable, distributed storage system. We are not limiting us to a particular system but appreciate anyone who would like to share about their experiences.

When: Thursday May 6th, 2010 at 6pm (open end)
Where: eCircle AG, Nymphenburger Straße 86, 80636 München ["Bruckmann" Building, "U1 Mailinger Str", map (in German) and look for the signs]

Thanks again to Bob Schulze from eCircle for providing the infrastructure.

We have a talk scheduled by Stefan Seelmann who is a member of the project committee for the Apache Directory project. This is followed by an open discussion.

Please RSVP at Xing and Yahoo's Upcoming.

Looking forward to seeing you there!

Cheers,
Lars

Let me take a minute to wrap up my FOSDEM 2010 experience. I was part of the NoSQL DevRoom organized by @stevenn from Outerthought, who I had the pleasure to visit before a few months back as an HBase Ambassador.

First things first, the NoSQL DevRoom was just an absolute success and I had a blast attending it. I also made sure to not walk around and see other talks outside the NoSQL track while there were many and plenty good ones. I did so deliberately to see the other projects and what they have to offer. I thought it was great, a good vibe was felt throughout the whole day as the audience got a whirlwind tour through the NoSQL landscape. The room was full to the brim for most presentations and some folks had to miss out as we could not have had more enter. This did prove the great interest in this fairly new kind of technology. Exciting!

The focus of my talk was about the history I have with HBase starting with it in late 2007. At this point I would like to take to the opportunity to thank Michael Stack, the lead of HBase, as he has helped me many times back then to sort out the problems I ran into. I would also like to say that if you start with HBase today you will not have these problems as HBase has matured tremendously since then. It is an overall stable product and can solve scalability issues you may face with regular RDBMS's today - and that with great ease.

So the talk I gave did not really sell all the features nor did it explain everything fully. I felt this could be left to the reader to look up on the project's website (or here on my blog) and hence I focused on my use case only. First up, here are the slides.


My Life with HBase - FOSDEM 2010 NoSQL -

After my talk and throughout the rest of the day I also had great conversations with the attendees who had many and great questions.

Having listened to the other talks though I felt I probably could have done a better job selling HBase to the audience. I could have reported about use-cases in well known companies, gave better performance numbers and so on. I have learned a lesson and am making sure I will be doing a better job next time around. I guess this is also another facet of what this is about, i.e. learning to achieve a higher level of professionalism.

But as I said above, my intend was to report about my life with HBase. I am grateful though that it was accepted as that and please let me cite Todd Hoff (see Hot Scalability Links for February 12, 2010) who put it in such nice words:

"The hardscabble tale of HBase's growth from infancy to maturity. A very good introduction and overview of HBase."

Thank you!

Finally here is the video of the talk:



I am looking forward to more NoSQL events in Europe in the near future and will attempt to represent HBase once more (including those adjustments I mentioned above). My hope is that we as Europeans are able to adopt these new technologies and stay abreast with the rest of the world. We sure have smart people to do so.

If you are staying on top of HBase development and frequently update to the HBase trunk (0.21.0-dev at the time of this post) you may have noticed that we now have support for Apache Ivy (see HBASE-1433). This is good because it allows to better control dependencies of the required jar files. It does have a few drawbacks though. One issue that you must be online to get your initial set of jars. You can also set up a local mirror or reuse the one you need for Hadoop anyways to share some of them.

Another issue is that it pulls in many more libs as part of the dependency resolving process. This reminds me bit of aptitude and when you try to install Java, for example on Debian. It often wants to pull in a litany of "required" packages but upon closer look many are only recommended and need not to be installed.

Finally you need to get the jar files somehow into your development environment. I am using Eclipse 3.5 on my Windows 7 PC as well as on my MacOS machines. If you have not run ant from the command line yet you have no jars downloaded and opening the project in Eclipse yields an endless amount of errors. You have two choices, you can run ant and get all jars and then add them to the project in Eclipse. But that is rather static and does not work well with future changes. It also is not the "ivy" way to resolve the libraries automatically.

The other option you have is adding a plugin to Eclipse that can handle Ivy for you, right within the IDE. Luckily for Eclipse there is IvyDE. You install it according to its documentation and then add a "Classpath Container" as described here. That part works quite well and after a restart IvyDE is ready to go.

A few more steps have to be done to get HBase working now - as in compiling without errors. The crucial one is editing the Ivy library and setting the HBase specific Ivy files. In particular the "Ivy settings path" and the properties file. The latter especially is specifying all the various version numbers that the ivy.xml is using. Without it the Ivy resolve process will fail with many errors all over the place. Please note that in the screen shot I added you see how it looks like on my Windows PC. The paths will be slightly different for your setup and probably even using another format if you are on a Mac or Linux machine. As long as you specify both you should be fine.

The other important issue is that you have to repeat that same step adding the Classpath Container two more times: each of the two larger contrib packages "contrib/stargate" and "contrib/transactional" have their own ivy.xml! For both you have to go into the respective directory and right click on the ivy.xml and follow the steps described in the Ivy documentation. Enter the same information as mentioned above to make the resolve work, leave everything else the way it is. You may notice that the contrib packages have a few more targets unticked. That is OK and can be used as-is.

As a temporary step you have to add two more static libraries that are in the $HBASE_HOME/lib directory: libthrift-0.2.0.jar and zookeeper-3.2.2.jar. Those will eventually be published on the Ivy repositories and then this step is obsolete (see INFRA-2461).

Eventually you end up with three containers as shown in the second and third screen shot. The Eclipse toolbar now also has an Ivy "Resolve All Dependencies" button which you can use to trigger the download process. Personally I had to do this a few times as the mirrors with the jars seem to be flaky at times. I ended up with for example "hadoop-mapred.jar" missing. Another resolve run fixed the problem.

The last screen shot shows the three Ivy related containers once more in the tree view of the Package Explorer in the Java perspective. What you also see is the Ivy console, which also is installed with the plugin. You have to open it as usual using the "Window - Show View - Console" menu (if you do not have the Console View open already) and then use the drop down menu next to the "Open Console" button in that view to open the Ivy console. It gives you access to all the details when resolving the dependencies and can hint when you have done something wrong. Please note though that it also lists a lot of connection errors, one for every mirror or repository that does not respond or yet has the required package available. One of them should respond though or as mentioned above you will have to try later again.

Eclipse automatically compiles the project and if everything worked out it does so now without a hitch. Good luck!

Update: Added info about the yet still static thrift and zookeeper jars. See Kay Kay's comment below.

What is the Write-ahead-Log you ask? In my previous post we had a look at the general storage architecture of HBase. One thing that was mentioned is the Write-ahead-Log, or WAL. This post explains how the log works in detail, but bear in mind that it describes the current version, which is 0.20.3. I will address the various plans to improve the log for 0.21 at the end of this article. For the term itself please read here.

Big Picture
The WAL is the lifeline that is needed when disaster strikes. Similar to a BIN log in MySQL it records all changes to the data. This is important in case something happens to the primary storage. So if the server crashes it can effectively replay that log to get everything up to where the server should have been just before the crash. It also means that if writing the record to the WAL fails the whole operation must be considered a failure.

Let"s look at the high level view of how this is done in HBase. First the client initiates an action that modifies data. This is currently a call to put(Put), delete(Delete) and incrementColumnValue() (abbreviated as "incr" here at times). Each of these modifications is wrapped into a KeyValue object instance and sent over the wire using RPC calls. The calls are (ideally batched) to the HRegionServer that serves the affected regions. Once it arrives the payload, the said KeyValue, is routed to the HRegion that is responsible for the affected row. The data is written to the WAL and then put into the MemStore of the actual Store that holds the record. And that also pretty much describes the write-path of HBase.

Eventually when the MemStore gets to a certain size or after a specific time the data is asynchronously persisted to the file system. In between that timeframe data is stored volatile in memory. And if the HRegionServer  hosting that memory crashes the data is lost... but for the existence of what is the topic of this post, the WAL!

We have a look now at the various classes or "wheels" working the magic of the WAL. First up is one of the main classes of this contraption.

HLog

The class which implements the WAL is called HLog. What you may have read in my previous post and is also illustrated above is that there is only one instance of the HLog class, which is one per HRegionServer. When a HRegion is instantiated the single HLog is passed on as a parameter to the constructor of HRegion.

Central part to HLog's functionality is the append() method, which internally eventually calls doWrite(). It is what is called when the above mentioned modification methods are invoked... or is it? One thing to note here is that for performance reasons there is an option for put(), delete(), and incrementColumnValue() to be called with an extra parameter set: setWriteToWAL(boolean). If you invoke this method while setting up for example a Put() instance then the writing to WAL is forfeited! That is also why the downward arrow in the big picture above is done with a dashed line to indicate the optional step. By default you certainly want the WAL, no doubt about that. But say you run a large bulk import MapReduce job that you can rerun at any time. You gain extra performance but need to take extra care that no data was lost during the import. The choice is yours.

Another important feature of the HLog is keeping track of the changes. This is done by using a "sequence number". It uses an AtomicLong internally to be thread-safe and is either starting out at zero - or at that last known number persisted to the file system. So as the region is opening its storage file, it reads the highest sequence number which is stored as a meta field in each HFile and sets the HLog sequence number to that value if it is higher than what has been recorded before. So at the end of opening all storage files the HLog is initialized to reflect where persisting has ended and where to continue. You will see in a minute where this is used.

The image to the right shows three different regions. Each of them covering a different row key range. As mentioned above each of these regions shares the the same single instance of HLog. What that means in this context is that the data as it arrives at each region it is written to the WAL in an unpredictable order. We will address this further below.

Finally the HLog has the facilities to recover and split a log left by a crashed HRegionServer. These are invoked by the HMaster before regions are deployed again.

HLogKey

Currently the WAL is using a Hadoop SequenceFile, which stores record as sets of key/values. For the WAL the value is simply the KeyValue sent from the client. The key is represented by an HLogKey instance. If you may recall from my first post in this series the KeyValue does only represent the row, column family, qualifier, timestamp, and value as well as the "Key Type". Last time I did not address that field since there was no context. Now we have one because the Key Type is what identifies what the KeyValue represents, a "put" or a "delete" (where there are a few more variations of the latter to express what is to be deleted, value, column family or a specific column).

What we are missing though is where the KeyValue belongs to, i.e. the region and the table name. That is stored in the HLogKey. What is also stored is the above sequence number. With each record that number is incremented to be able to keep a sequential order of edits. Finally it records the "Write Time", a time stamp to record when the edit was written to the log.

LogFlusher

As mentioned above as data arrives at a HRegionServer in form of KeyValue instances it is written (optionally) to the WAL. And as mentioned as well it is then written to a SequenceFile. While this seems trivial, it is not. One of the base classes in Java IO is the Stream. Especially streams writing to a file system are often buffered to improve performance as the OS is much faster writing data in batches, or blocks. If you write records separately IO throughput would be really bad. But in the context of the WAL this is causing a gap where data is supposedly written to disk but in reality it is in limbo. To mitigate the issue the underlaying stream needs to be flushed on a regular basis. This functionality is provided by the LogFlusher class and thread. It simply calls HLog.optionalSync(), which checks if the  hbase.regionserver.optionallogflushinterval, set to 10 seconds by default, has been exceeded and if that is the case invokes HLog.sync(). The other place invoking the sync method is HLog.doWrite(). Once it has written the current edit to the stream it checks if the hbase.regionserver.flushlogentries parameter, set to 100 by default, has been exceeded and calls sync as well.

Sync itself invokes HLog.Writer.sync() and is implemented in SequenceFileLogWriter. For now we assume it flushes the stream to disk and all is well. That in reality this is all a bit more complicated is discussed below.

LogRoller

Obviously it makes sense to have some size restrictions related to the logs written. Also we want to make sure a log is persisted on a regular basis. This is done by the LogRoller class and thread. It is controlled by the hbase.regionserver.logroll.period parameter in the $HBASE_HOME/conf/hbase-site.xml file. By default this is set to 1 hour. So every 60 minutes the log is closed and a new one started. Over time we are gathering that way a bunch of log files that need to be maintained as well. The HLog.rollWriter() method, which is called by the LogRoller to do the above rolling of the current log file, is taking care of that as well by calling HLog.cleanOldLogs() subsequently. It checks what the highest sequence number written to a storage file is, because up to that number all edits are persisted. It then checks if there is a log left that has edits all less than that number. If that is the case it deletes said logs and leaves just those that are still needed.


The other parameters controlling the log rolling are hbase.regionserver.hlog.blocksize and hbase.regionserver.logroll.multiplier, which are set by default to rotate logs when they are at 95% of the blocksize of the SequenceFile, typically 64M. So either the logs are considered full or when a certain amount of time has passed causes the logs to be switched out, whatever comes first.

Replay

Once a HRegionServer starts and is opening the regions it hosts it checks if there are some left over log files and applies those all the way down in Store.doReconstructionLog(). Replaying a log is simply done by reading the log and adding the contained edits to the current MemStore. At the end an explicit flush of the MemStore (note, this is not the flush of the log!) helps writing those changes out to disk.

The old logs usually come from a previous region server crash. When the HMaster is started or detects that region server has crashed it splits the log files belonging to that server into separate files and stores those in the region directories on the file system they belong to. After that the above mechanism takes care of replaying the logs. One thing to note is that regions from a crashed server can only be redeployed if the logs have been split and copied. Splitting itself is done in HLog.splitLog(). The old log is read into memory in the main thread (means single threaded) and then using a pool of threads written to all region directories, one thread for each region.

Issues

As mentioned above all edits are written to one HLog per HRegionServer. You would ask why that is the case? Why not write all edits for a specific region into its own log file? Let's quote the BigTable paper once more:

If we kept the commit log for each tablet in a separate log file, a very large number of files would be written concurrently in GFS. Depending on the underlying file system implementation on each GFS server, these writes could cause a large number of disk seeks to write to the different physical log files.

HBase followed that principle for pretty much the same reasons. As explained above you end up with many files since logs are rolled and kept until they are safe to be deleted. If you do this for every region separately this would not scale well - or at least be an itch that sooner or later is causing pain.

So far that seems to be no issue. But again, it causes problems when things go wrong. As long as you have applied all edits in time and persisted the data safely, all is well. But if you have to split the log because of a server crash then you need to divide into suitable pieces, as described above in the "replay" paragraph. But as you have seen above as well all edits are intermingled in the log and there is no index of what is stored at all. For that reason the HMaster cannot redeploy any region from a crashed server until it has split the logs for that very server. And that can be quite a number if the server was behind applying the edits.

Another problem is data safety. You want to be able to rely on the system to save all your data, no matter what newfangled algorithms are employed behind the scenes. As far as HBase and the log is concerned you can turn down the log flush times to as low as you want - you are still dependent on the underlaying file system as mentioned above; the stream used to store the data is flushed but is it written to disk yet? We are talking about fsync style issues. Now for HBase we are most likely talking Hadoop's HDFS as being the file system that is persisted to.

Up to this point it should be abundantly clear that the log is what keeps data safe. For that reason a log could be kept open for up to an hour (or more if configured so). As data arrives a new key/value pair is written to the SequenceFile and occasionally flushed to disk. But that is not how Hadoop was set out to work. It was meant to provide an API that allows to open a file, write data into it (preferably a lot) and closed right away, leaving an immutable file for everyone else to read many times. Only after a file is closed it is visible and readable to others. If a process dies while writing the data the file is pretty much considered lost. What is required is a feature that allows to read the log up to the point where the crashed server has written it (or as close as possible).


While append for HDFS in general is useful it is not used in HBase, but the hflush() is. What it does is writing out everything to disk as the log is written. In case of a server crash we can safely read that "dirty" file up to the last edits. The append in Hadoop 0.19.0 was so badly suited that a hadoop fsck / would report the DFS being corrupt because of the open log files HBase kept.

Bottom line is, without Hadoop 0.21.0 you can very well face data loss. With Hadoop 0.21.0 you have a state-of-the-art system.

Planned Improvements

For HBase 0.21.0 there are quite a few things lined up that affect the WAL architecture. Here are some of the noteworthy ones.

SequenceFile Replacement

One of the central building blocks around the WAL is the actual storage file format. The used SequenceFile has quite a few shortcomings that need to be addressed. One for example is the suboptimal performance as all writing in SequenceFile is synchronized, as documented in HBASE-2105.

As with HFile replacing MapFile in HBase 0.20.0 it makes sense to think about a complete replacement. A first step was done to make the HBase classes independent of the underlaying file format. HBASE-2059 made the class implementing the log configurable.

Another idea is to change to a different serialization altogether. HBASE-2055 proposes such a format using Hadoop's Avro as the low level system. Avro is also slated to be the new RPC format for Hadoop, which does help as more people are familiar with it.

Append/Sync

Even with hflush() we have a problem that calling it too often may cause the system to slow down. Previous tests using the older syncFs() call did show that calling it for every record slows down the system considerably. One step to help is to implement a "Group Commit", done in HBASE-1939. It flushes out records in batches. In addition HBASE-1944 adds the notion of a "deferred log flush" as a parameter of a Column Family. If set to true it leaves the syncing of changes to the log to the newly added LogSyncer class and thread. Finally HBASE-2041 sets the flushlogentries to 1 and optionallogflushinterval to 1000 msecs. The .META. is always synced for every change, user tables can be configured as needed.

Distributed Log Splitting

As remarked splitting the log is an issue when regions need to be redeployed. One idea is to keep a list of regions with edits in Zookeeper. That way at least all "clean" regions can be deployed instantly. Only those with edits need to wait then until the logs are split.

What is left is to improve how the logs are split to make the process faster. Here is how is the BigTable addresses the issue:

One approach would be for each new tablet server to read this full commit log file and apply just the entries needed for the tablets it needs to recover. However, under such a scheme, if 100 machines were each assigned a single tablet from a failed tablet server, then the log file would be read 100 times (once by each server).

and further

We avoid duplicating log reads by first sorting the commit log entries in order of the keys (table, row name, log sequence number). In the sorted output, all mutations for a particular tablet are contiguous and can therefore be read efficiently with one disk seek followed by a sequential read. To parallelize the sorting, we partition the log file into 64 MB segments, and sort each segment in parallel on different tablet servers. This sorting process is coordinated by the master and is initiated when a tablet server indicates that it needs to recover mutations from some commit log file.

This is where its at. As part of the HMaster rewrite (see HBASE-1816) the log splitting will be addressed as well. HBASE-1364 wraps the splitting of logs into one issue. But I am sure that will evolve in more sub tasks as the details get discussed.

At the end of last year we had a first meeting in Munich to allow everyone interested in all things Hadoop and related technologies to gather, get to know each other and exchange their knowledge, experiences and information. We would like to invite for the second Munich OpenHUG Meeting and hope to see you all again and meet even more new enthusiasts there. We would also be thrilled if those attending would be willing to share their story so that we can learn about other projects and how people are using the exciting NoSQL related technologies. No pressure though, come along and simply listen if you prefer, we welcome anyone and everybody!

When: Thursday February 25, 2010 at 5:30pm open end
Where: eCircle AG, Nymphenburger Straße 86, 80636 München ["Bruckmann" Building, "U1 Mailinger Str", map (in German) and look for the signs]

Thanks again to Bob Schulze from eCircle to provide the location and projector. So far we have a talk scheduled by Christoph Rupp about HamsterDB. We are still looking for volunteers who would like to present on any related topic (please contact me)! Otherwise we will have an open discussion about whatever is brought up by the attendees.

Last but not least there will be something to drink and we will get pizzas in. Since we do not know how many of you will come we simply stay at the events location and continue our chats over food.

Looking forward to seeing you there!

Please RSVP at Upcoming or Xing.