Author: skahler

I was reading an excellent post ( When should I use Greenplum Database versus HAWQ? ) by my colleague Jon Roberts. Which got me thinking I should drop a post on why I still think MPP database and specifically Greenplum DB are relevant in a Hadoop crazy world. It should be noted that I am absolutely a Hadoop fanboy. I have managed to make my way to the last three Strata conferences and received my CDH certification before they had a pretty manager to take care of everything, so I’m no hater.

In considering Hadoop there are a few key items you should consider:

• Big Data ≠ Hadoop
• Hadoop is a platform

I’ll tackle the last first, Hadoop is a platform. Hadoop is made up a variety of components of which various sources will define in different ways. What all of them boil down to is a distributed datastore (most often HDFS ) with distributed processing implementation on top of it (most often MapReduce). This is changing landscape and often multiple implementations of processing are being offered in order to bring different capabilities forward. The net of this is much like implementing virtualization, you are going to need someone skilled in Hadoop to bring out it’s value and find the right use cases. In addition you also need someone who understands Hadoop infrastructure, it is fundamentally different than the standard infrastructure most companies have been moving to. Probably the biggest hurdle will be that you need to convince business units to align to doing things in a potentially new way on this new platform. The way in which they query, store and even load data will most likely change.  While there are a variety of tools and vendors out there poised to help ease and support this transition, realize this is something new on the scale of implementing virtualization, SAN storage or shifting to cloud hosting. Nobody will debate value can be found in that list of technologies, as long as you are willing to find the right use case and then bite the bullet and implement them. Much the same with Hadoop.

There is also a common misconception out there that Hadoop is the only path to work with Big Data. While a majority of the press revolves around what Google, Yahoo and Facebook are doing, the fact of the matter is that most companies Big Data is not at that scale. Many products exist out there that scale in a distributed fashion to handle datasets that span well beyond one rack or multiple racks without needing to implement the Hadoop platform beneath it. Standard infrastructure skills can be utilized with a much smaller learning curve to get results from the implementation of these technologies. Often these products exhibit greater analytics ability and/or faster data processing capabilities. What has caused Hadoop to thrive is not necessarily what kind of processing Hadoop can do in many cases. It’s that it is perceived as an alternative to do a subset of what is currently being done at a cheaper cost and people are hacking to add that missing functionality. While cost is an extremely important factor, if not the one important factor, it shows that Big Data processing is not exclusive to Hadoop. Just that what Hadoop has done for many companies is make Big Data look more palatable.

This is leads me to my view on the relevance of MPP databases. I believe over the next year as products like Hive and Impala permeate the market, things like HAWQ and Presto become more widely known and pieces like Stinger and Tez drive the speed of SQL on Hadoop it will push down the cost of the MPP database and thus the reason many companies are looking to Hadoop as an alternative. In addition people considering the move are going to get a more realistic look at what other companies have run into as they move Hadoop to production and what the true cost of bringing that infrastructure up are. As the delta between MPP and the true cost of Hadoop shrink you will see less positioning of Hadoop as the rip and replace for the MPP database but as an augmenter that sits along side of it and can be leveraged to do bring value to the right use cases.

In short, Hadoop is the new kid on the block and it won’t replace the MPP database. You can plan on there being a few fights though before they become good friends.

Greenplum Database 4.3 rolled out earlier this month and there are some big changes.

• Append-Optimized Tables: UPDATE and DELETE operations are enabled on what were previously Append-Only tables.
• Master Mirroring Enhanced: The way Master/StandbyMaster replication is implemented has changed. This should deliver faster master failover and many more of the functions related to it can now be done online.

Docs added to the docs page here

Data nerd goodness from the 451 group

Since the roll out of Greenplum 4.2 in combination with Greenplum on-site for a little training our clusters stability over the last year has improved. The main issues we run into at this point in time is lack of space as people want to throw more and more on the clusters and the occasional bad query that gets into the system and causes issues. To help with some of the space issues some simple jobs have been moved to a Hadoop cluster, this allows us to use Greenplum for the more complex data analytics functions. By removing some of the rollup functionality we have cut a decent number of table scans out of the system which resulted in a definite responsiveness increase.

The other problem we run into is queries getting into the system that cause resource contention. We have a liberal access policy and as it’s a startup environment, change is constant. Tools thus far to track down a rogue query haven’t been exactly outstanding. With the acquisition of MoreVRP I’m hoping a blessed and bottled better solution for this will come out. Currently I gather process stats and drop them to a local filesystem and drag that into a local share to do pivot tables on it and look for issues. There are some in database Greenplum queries that do this, of course when the DB is having issues using it to troubleshoot why it is having issues doesn’t work very well. The past couple of weeks we’ve been doing some work with OpenTSDB and it looks very promising. The tcollector system is easy to work with and the amount of data we have been able to throw at a single TSDB instance is impressive. The display ability of OpenTSDB is it’s weak point, we haven’t found a good display layer. Using the current tools though I am able to product the following graph, which shows memory utilization by query on each segment. Stats are collected every minute so a lot of queries don’t show, but our sub-minute queries aren’t really our problem children. As I perfect the filters I think this will be extremely helpful in tracking down skewing queries and the memory hogs the get dropped on the system in realtime.

One of our clusters we tried out used C2100s and these LSI 9260-8i controllers. They ran fantastically out of the gate, but we started to run into some issues. I can attest to the fact that our rack of servers ran hot. In pulling an over nighter it would get cold in the facility and I’d stand behind that rack because the heat coming out of over a dozen servers with a dozen disks running and processors running full bore kept my fingers from cramping up. This wasn’t heat senor warning hot, but it definitely put out some warmth. Later on when we started to have issues we’d pull the controller out and find the heat sink sitting on top of the card not attached in any way. What I think the problem was is that these controllers have the heat sink held on by plastic clips and springs and as our servers would run warm, not to the sensor heat warning level, eventually the clips would melt leaving the heatsink sitting there on top of the controller.