Apache HBase Do’s and Don’ts

I recently gave a talk at the LA Hadoop User Group about Apache HBase Do’s and Don’ts. The audience was excellent and had very informed and well articulated questions. Jody from Shopzilla was an excellent host and I owe him a big thanks for giving the opportunity to speak with over 60 LA Hadoopers. Since not everyone lives in LA or could make it to the meetup, I’ve summarized some of the salient points here. For those of you with a busy day, here’s the tl;dr:

  • HBase is good, but not an RDBMS or HDFS replacement
  • Good configuration means good operation
  • Monitor monitor monitor monitor monitor

We at Cloudera are big fans of HBase. We love the technology, we love the community and we’ve found that it’s a great fit for many applications. Successful uses of HBase have been well documented and as a result, many organizations are considering whether HBase is a good fit for some of their applications. The impetus for my talk and this follow up blog post is to clarify some of the good applications for HBase, warn against some poor applications and highlight important steps to a successful HBase deployment.

When to use HBase

The most important consideration when looking at HBase is that, while it is a great solution to many problems, it is not a silver bullet. HBase is not optimized for classic transactional applications or even relational analytics. It is also not a complete substitute for HDFS when doing large batch MapReduce. Take a look at some of the use cases in this post to get a sense of which applications are a good fit for HBase and if you have questions, go ahead and post on the lists. Have I mentioned that the community is fantastic?

With that caveat out the way – why should you use HBase? If your application has a variable schema where each row is slightly different, then you should look at HBase. As an example, doing a modeling exercise using a standard relational schema; When you can’t add columns fast enough and most of them are NULL in each row, you should consider HBase. If you find that your data is stored in collections, for example some meta data, message data or binary data that is all keyed on the same value, then you should consider HBase. If you need key based access to data when storing or retrieving, then you should consider HBase.

Supporting Services

Assuming you’re convinced that HBase is a good fit for your application, here are some tips you need to consider when deploying it. There are a few supporting services that are important and one that’s required. If you haven’t looked at ZooKeeper before, now is the time. HBase uses ZooKeeper for various distributed coordination services such as master election. As HBase develops and grows it continues to rely on ZooKeeper for additional functionality, making it a key part of the system. In addition, you should have proper network services in place such as NTP and DNS. HBase depends on all nodes in the cluster having closely synchronized clocks and referring to each other consistently. Using NTP and DNS ensures that you won’t run into odd behaviors when one node A thinks that the time is tomorrow and node B thinks it’s yesterday. You’ll also prevent situations where the master node tells node C to serve a region but node C doesn’t know its own name and doesn’t answer. Using NTP and DNS will save a lot of headaches as you get started.

I’ve said that the most important consideration when selecting HBase is to make sure you have a use case that fits. The most important thing to do when using HBase is to monitor the system. Monitoring is key to successful HBase operations. As is the case with many distributed systems, HBase is susceptible to cascading failures. If one node starts swapping it can lose contact with the master, causing another sever to pick up the load and becoming over burdened. That second server will fail and the failure will cascade. You need to monitor the memory, the CPU, the I/O and the network latency and bandwidth on each of your HBase nodes to make sure they are operating within healthy parameters. Monitoring is the most important practice to operating a healthy HBase cluster.

Good Practices for HBase Architecture

Fast forward to your well-monitored HBase cluster running a perfect use case, here are some good practices. Use a key prefix that distributes well based on your use case. If you prefix your key by timestamp or any similar value that, when sorted, is stored or queried in a batch then you will likely overload each region server in turn instead of evenly distributing the load. You should also keep the number of regions to a reasonable number based on memstore size and amount of RAM and the RegionServer JVM should be limited to 12GB of java heap to minimize long GC pauses. For example a machine with 36GB of RAM that is also running a DataNode daemon could handle approximately 100 regions with active writes and a memstore of 48MB each. That allows enough headroom for DataNode and RegionServer memory requirements, Linux file buffer space and a reasonable flush size for each RegionServer.

A few configuration recommendations include disabling auto-compaction (by default it happens every 24 hours from the time you start HBase) and schedule it to run every day at an off-peak time. You should also configure compression (such as LZO) and explicitly put the correctly configured HBase conf directory in your CLASSPATH.

HBase DON’Ts

We’ve covered a broad range of good practices for HBase. There are also a few use patterns to avoid. For example, don’t expect to use HBase as a wholesale replacement for every one of your relational databases. HBase is great at many things but it doesn’t replace relational databases. For a start, it doesn’t talk SQL, have an optimizer, support cross record transactions or joins. If you don’t use any of these in your database application then HBase could very well be the perfect fit.

Be careful when running mixed workloads on an HBase cluster. When you have SLAs on HBase access independent of any MapReduce jobs (for example, a transformation in Pig and serving data from HBase) run them on separate clusters. HBase is CPU and Memory intensive with sporadic large sequential I/O access while MapReduce jobs are primarily I/O bound with fixed memory and sporadic CPU. Combined these can lead to unpredictable latencies for HBase and CPU contention between the two. A shared cluster also requires fewer task slots per node to accommodate for HBase CPU requirements (generally half the slots on each node that you would allocate without HBase). Also keep an eye on memory swap. If HBase starts to swap there is a good chance it will miss a heartbeat and get dropped from the cluster. On a busy cluster this may overload another region, causing it to swap and a cascade of failures.

Final Thoughts

One last bit of advice before we sum up. When loading HBase, use HFileOuputFormat if loading via a MapReduce job or a collection of servers using batched puts. Loading with a single client will bottleneck on that client and not take advantage of the scalability afforded by HBase.

In summary, consider HBase when you’re loading data by key, searching data by key (or range), serving data by key, querying data by key or when storing data by row that doesn’t conform well to a schema.


Use Cases

Filed under:

9 Responses
  • jimmy / April 20, 2011 / 2:55 PM

    change the hadoop-0.20-tasktracker so the process is started like this:

    daemon nice -n 19 ionice -c2 -n7 /usr/lib/hadoop-0.20/bin/hadoop-daemon.sh –
    config “/etc/hadoop-0.20/conf” start tasktracker $DAEMON_FLAGS

    helps to solve the IO contention issue when hbase regionserver and hadoop datanode coexist on the same physical machine. it lower the priority of the mapreduce task’s IO. Another configuration we should do is to minimize the configured map/reduce slots so it won’t compete too aggressively with hbase.

  • Vlad / April 21, 2011 / 2:38 PM

    “You need to monitor the memory, the CPU, the I/O and the network latency and bandwidth on each of your HBase nodes to make sure they are operating within healthy parameters”

    You forgot to mention ” and pray that something won’t break and goes down the very next moment”

    Amen.

  • Otis Gospodnetic / April 27, 2011 / 10:49 PM

    Is there a good HBase monitoring software or service that you would recommend, other than Ganglia?

  • william johnson / May 26, 2011 / 1:26 PM

    i agree that HBase is good, but not an RDBMS or HDFS replacement…………….

  • Otis Gospodnetic / March 10, 2012 / 12:53 PM

    Heh, Q from almost 1 year ago:

    > Is there a good HBase monitoring software or
    > service that you would recommend, other than
    > Ganglia?

    At this point we use our own SPM for HBase to monitor HBase performance. Other HBase users may find it useful, too: http://sematext.com/spm/hbase-performance-monitoring/index.html

Leave a comment


5 × = twenty five