NoSQL comparison: All content tagged as NoSQL comparison in NoSQL databases and polyglot persistence
Every decent benchmark can reveal not only performance or stability problems, but oftentimes more subtle issues like less known or undocumented options, common misconfigurations or misunderstandings. Sometimes it can reveal scenarios that a product hasn’t considered before or for which it has different solutions.
What can be learned from the Hypertable vs HBase benchmark from the HBase perspective?
Michael Stack: That we need to work on our usability; even a smart fellow like Doug Judd can get it really wrong.
We haven’t done his sustained upload in a good while. Our defaults need some tweaking.
We need to do more documentation around JVM tuning; you’d think fellas would have grok’d by now that big java apps need their JVM’s tweaked but it looks like the message still hasn’t gotten out there.
That we need a well-funded PR dept. to work on responses to the likes of Doug’s article (well-funded because Doug claims he spent four months on his comparison).
Jean-Daniel Cryans: I already opened a few jiras after using HT’s test on a cluster I have here with almost the same hardware and node count, it’s mostly about usability and performance for that type of use case:
Hypertable does a neat thing where it changes the size given to the CellCache (our MemStores) and Block Cache based on the workload. If you need an image, scroll down at the bottom of this link:
Coming out of HBASE-2375, we need a new functionality much like hypertable’s where we would have a lower split size for new tables and it would grow up to a certain hard limit. This helps usability in different ways:
- With that we can set the default split size much higher and users will still have good data distribution
- No more messing with force splits
- Not mandatory to pre-split your table in order to get good out of the box performance
The way Doug Judd described how it works for them, they start with a low value and then double it every time it splits. For example if we started with a soft size of 32MB and a hard size of 2GB, it wouldn’t be until you have 64 regions that you hit the ceiling.
On the implementation side, we could add a new qualifier in .META. that has that soft limit. When that field doesn’t exist, this feature doesn’t kick in. It would be written by the region servers after a split and by the master when the table is created with 1 region.
Spawning this from HBASE-2375, I saw that it was much more efficient compaction-wise to check if we can split right after flushing. Much like the ideas that Jon spelled out in the description of that jira, the window is smaller because you don’t have to compact and then split right away to only compact again when the daughters open.
If someone is faced with similar scenarios are there workarounds or different solutions?
Michael Stack: There are tunings of HBase configs over in our reference guide for the sustained upload both in hbase and in jvm.
Then there is our bulk load facility which by-passes this scenario altogether which is what we’d encourage folks to use because its 10x to 100x faster getting your data in there.
Jean-Daniel Cryans: You can import 5TB in HBase with sane configs, I’ve done it a few times already since I started using his test. The second time he ran his test he just fixed mslab but still kept the crazy ass other settings like 80% of the memory dedicated to memstores. My testing also shows that you need to keep the eden space under control, 64MB seems a good value in my testing (he didn’t set any in his test, the first time I ran mine without setting it I got the concurrent mode failure too).
The answer he gave this week to Todd’s email on the hadoop mailing list is about a constant stream of updates and that’s what he’s trying to test. Considering that the test imports 5TB in ~16h (on my cluster), you run out of disk space in about 3 days. I seriously don’t know what he’s aiming for here.
Quoting him: “Bulk loading isn’t always an option when data is streaming in from a live application. Many big data use cases involve massive amounts of smaller items in the size range of 10-100 bytes, for example URLs, sensor readings, genome sequence reads, network traffic logs, etc.”
What are the most common places to look for improving the performance of a HBase cluster?
Michael Stack: This is what we point folks at when they ask the likes of the above question: HBase Performance Tunning
If that chapter doesn’t have it, its a bug and we need to fix up our documentation more.
Jean-Daniel Cryans: What Stack said. Also if you run into GC issues like he did then you’re doing it wrong.
I think HBase should find all of this challenging and flattering. Challenging because we know how we can do better along the dimensions of your testing and you are kicking us pretty hard. Flattering because by inference we seem to be worth kicking.
But this misses the point, and reduces what should be a serious discussion of the tradeoffs between Java and C++ to a cariacture. Furthermore, nobody sells HBase. (Not in the Hypertable or Datastax sense. Commercial companies bundle HBase but they do so by including a totally free and zero cost software distribution.) Instead it is voluntarily chosen for hundreds of large installations all over the world, some of them built and run by the smartest guys I have ever encountered in my life. Hypertable would have us believe we are all making foolish choices. While it is true that we all on some level have to deal with the Java heap, only Hypertable seems to not be able to make it work. I find that unsurprising. After all, until you can find some way to break it, you don’t have any kind of marketing story.
This remineded me of the quote from Jonathan Ellis’s Dealing With JVM Limitations in Apache Cassandra:
Cliff Click: Many concurrent algorithms are very easy to write with a GC and totally hard (to down right impossible) using explicit free.
As I was expecting, there are quite a few good things that will come out from this benchmark for both long time HBase users, but also for new adopters.
Original title and link: What HBase Learned From the Hypertable vs HBase Benchmark ( ©myNoSQL)
A great post by Olivier Mallassi on a topic that comes up very often: how do data grids and NoSQL databases compare?
- Data Grids enable you controlling the way data is stored. They all have default implementation (Gigaspaces offers RDBMS by default, Gemfire offers file and disk based storage by default….) but in all cases, you can choose the one that fits your needs: do you need to store data, do you need to relieve the existing databases….
- In order to minimize the latency, data grids enable you to store data synchronously (write-through) or asynchronously (write-behind) on disk. You can also define overflow strategies. In that case, data is store in memory up to a treshold where data is flushed on disk (following algorithms like LRU …). NoSQL solutions have not been designed to provide these features.
- Data grids enable you developing Event Driven Architecture.
- Querying is maybe the point on which pure NoSQL solutions and data grids are merging.
- Data grids enable near-cache topologies.
Taking a step back you’ll notice that there are actually more similarities than differences. While Oliver Mallasi lists the above points as features that prove data grids as being more configurable and so more adaptable, some of these do exist also in the NoSQL databases taking different forms:
- pluggable storage backends. Not many of the NoSQL databases have this feature,but Riak and Project Voldemort are offering different solutions that are optimized for specific scenarios.
- replicated and durable writes. Not the same as synchronous vs asynchronous writes, but a different perspective on writes.
- Notification mechanisms. Once again not all of the NoSQL databases support notification mechanisms, but a couple of them have offer some interesting approaches:
- Most of the NoSQL database have local per-node caches.
With these, I’ve probably made things even blurrier. But let me try to draw a line between data grids and NoSQL databases. Data grids are optimized for handling data in memory. Everything that spills over is secondary. On the other hand, NoSQL databases are for storing data. Thus they focus on how they organize data (on disk or in memory) and optimize access to it. Data grids are a processing/architectural model. NoSQL databases are storage solutions.
Original title and link: Data Grid or NoSQL? What are the common points? The main differences? ( ©myNoSQL)
Though it looks like mongo-store demonstrates the best overall performance, it should be noted that a mongo server is unlikely to be used solely for caching (the same applies to redis), it is likely that non-caching related queries will be running concurrently on a mongo/redis server which could affect the suitability of these benchkmarks.
I’m not a Rails user, so please take these with a grain of salt:
without knowing the size of the cached objects, at 20000 iterations most probably neither MongoDB, nor Redis have had to persist to disk.
This means that all three of memcached, MongoDB, Redis stored data in memory only
if no custom object serialization is used by any of the memcached, MongoDB, Redis caches, then the performance difference is mostly caused by the performance of the driver
it should not be a surprise to anyone that the size of the cached objects can and will influence the results of such benchmarks
there doesn’t seem to be any concurrent access to caches. Concurrent access and concurrent updates of caches are real-life scenarios and not including them in a benchmark greatly reduces the value of the results
none of these benchmarks doesn’t seem to contain code that measure the performance of cache eviction
Except the case where any of these forces a disk write ↩
Original title and link: Rails Caching Benchmarked: MongoDB, Redis, Memcached ( ©myNoSQL)
InterSystems, producers of the Caché database, have launched Globals, a fast, proven, simple, flexible and free databases, 2 months ago. But after the initial announcement, I couldn’t find and didn’t hear much about it. This until Rob Tweed and K.S.Bhaskar took the time to explained some of the differences between InterSystems Globals and GT.M, both systems being implemented on top of the MUMPS Global Persistent Variables .
Rob Tweed: I’m not an InterSystems person — simply a long-term user and advocate of Global-storage based technologies of which GT.M, Cache and now InterSystem Globals are members, and someone who has long believed that it’s a significantly under-valued database technology, and unfortunately and sadly not known about or understood sufficiently in the wider database/IT world. However, the rise of NoSQL has provided some renewed chance of rediscovery by a wider community of developers, which I’m keen to encourage.
With respect to a comparison with BigTable etc, I guess all of us in the Global-storage technology user communities have looked at many of the new NoSQL technologies and thought it’s deja vu all over again :-) Perhaps this paper that I co-authored might help to at least provide a comparative positioning against the “mainstream” NoSQL databases.
As we note in the paper, full-blown Cache and GT.M provide many of the mechanisms needed for high-end scalability, though, as you point out, many of these appear to be lacking in InterSystem Globals, at least in its current (and relatively early) incarnation.
Regarding a comparison of InterSystem Globals and GT.M, at the core data storage level, there’s little difference: they both use Globals for data storage, so the use cases will be similar. Both GT.M and Globals are implemented in C (instead of M/Mumps), with some small bits of GT.M glue code in assembler. In terms of licensing, InterSystem Globals is free but proprietary, GT.M is free open source.
I guess the biggest differences are:
InterSystem Globals is essentially the core database engine from Cache, but with many of the features of Cache, in particular its native language (M) turned off. The concept in InterSystem Globals is that it will be accessed via APIs from other mainstream languages, instead of being primarily accessed via the M language as is the norm in, say, GT.M
K.S.Bhaskar: Although the majority of GT.M users do indeed program in M, the fact is that the GT.M database is just as accessible from a C
(KSB) As discussed above, GT.M does not restrict a user to TCP access. The primary restriction (which results from the fact that the database engine is daemonless and processes cooperate to manage the database - so there is a real time database engine linked into each processes’ address space) is that a GT.M process can have only one thread. If you can’t live with this, then you have to use TCP through a client such as Rob’s.
Another option is to use the GT.CM “database service” that GT.M includes (GNP - the GT.M Network Protocol is layered on TCP). A client is coded within GT.M itself, or you can use/adapt other clients for GNP such as Dave H’s PHP gtcmclient.
(RT): I suspect one way things will pan out over coming months will be:
- If you want the ultimate in performance and willing to sacrifice open source and the high-end scalability options, but remain free, then InterSystem Globals will be a good choice
If you want the former but are willing to pay for the extra high-end scalability technologies, then full-blown Cache will be your choice.
(KSB) This is a false choice. GT.M gives you high end scalability with a free / open source license (and support with assured service levels on commercial terms for those who want it).
(RT) The nice thing is that it will be straightforward to engineer applications that can be easily migrated between these three options with a minimum of change being needed at the application level.
Personally, I think InterSystem Globals is a great thing and nice to see InterSystems venturing into a new direction: I think that’s only to be encouraged and can only help the NoSQL community.
The text of this post has been adapted and edited based on this conversation .