Analysis of implementations of and issues associated with the parallel programming framework MapReduce. Related subjects include:
I talked tonight with Lee Edlefsen, Chief Scientist of Revolution Analytics, and now think I understand Revolution’s parallel R much better than I did before.
There are four primary ways that people try to parallelize predictive modeling:
- They can run the same algorithm on different parts of a dataset on different nodes, then return all the results, and claim they’ve parallelized. This is trivial and not really a solution. It is also the last-ditch fallback position for those who parallelize more seriously.
- They can generate intermediate results from different parts of a dataset on different nodes, then generate and return a single final result. This is what Revolution does.
- They can parallelize the linear algebra that underlies so many algorithms. Netezza and Greenplum tried this, but I don’t think it worked out very well in either case. Lee cited a saying in statistical computing “If you’re using matrices, you’re doing it wrong”; he thinks shortcuts and workarounds are almost always the better way to go.
- They can jack up the speed of inter-node communication, perhaps via MPI (Messaging Passing Interface), so that full parallelization isn’t needed. That’s SAS’ main approach.
One confusing aspect of this discussion is that it could reference several heavily-overlapping but not identical categories of algorithms, including:
- External memory algorithms, which operates on datasets too big to fit in main memory, by — for starters — reading in and working on a part of the data at a time. Lee observes that these are almost always parallelizable.
- What Revolution markets as External Memory Algorithms, which are those external memory algorithms it has gotten around to implementing so far. These are all parallelized. They are also all in the category of …
- … algorithms that can be parallelized by:
- Operating on data in parts.
- Getting intermediate results.
- Combining them in some way for a final result.
- Algorithms of the previous category, where the way of combining them specifically is in the form of summation, such as those discussed in the famous paper Map-Reduce for Machine Learning on Multicore. Not all of Revolution’s current parallel algorithms fall into this group.
To be clear, all Revolution’s parallel algorithms are in Category #2 by definition and Category #3 in practice. However, they aren’t all in Category #4.
|Categories: Greenplum, Hadoop, MapReduce, Netezza, Parallelization, Predictive modeling and advanced analytics, Revolution Analytics, Teradata||Leave a Comment|
When we scheduled a call to talk about Sentry, Cloudera’s Charles Zedlewski and I found time to discuss other stuff as well. One interesting part of our discussion was around the processing “frameworks” Cloudera sees as most important.
- The four biggies are:
- MapReduce. Duh.
- SQL, specifically Impala. This is as opposed to the uneasy Hive/MapReduce layering.
- “Math” , which seems to mainly be through partnerships with SAS and Revolution Analytics. I don’t know a lot about how these work, but I presume they bypass MapReduce, in which case I could imagine them greatly outperforming Mahout.
- Stream processing (Storm) is next in line.
- Graph — e.g. Giraph — rises to at least the proof-of-concept level. Again, the hope would be that this well outperforms graph-on-MapReduce.
- Charles is also seeing at least POC interest in Spark.
- But MPI (Message Passing Interface) on Hadoop isn’t going anywhere fast, except to the extent it’s baked into SAS or other “math” frameworks. Generic MPI use cases evidently turn out to be a bad fit for Hadoop, due to factors such as:
- Low data volumes.
- Latencies in various parts of the system
HBase was artificially omitted from this “frameworks” discussion because Cloudera sees it as a little bit more of a “storage” system than a processing one.
Another good subject was offloading work to Hadoop, in a couple different senses of “offload”: Read more
|Categories: Cloudera, Complex event processing (CEP), Databricks, Spark and BDAS, Endeca, Hadoop, HP and Neoview, MapReduce, Predictive modeling and advanced analytics, RDF and graphs, Revolution Analytics, SAS Institute, Teradata||22 Comments|
I made a remarkably rumpled video appearance yesterday with SiliconAngle honchos John Furrier and Dave Vellante. (Excuses include <3 hours sleep, and then a scrambling reaction to a schedule change.) Topics covered included, with approximate timechecks:
- 0:00 Introductory pabulum, and some technical difficulties
- 2:00 More introduction
- 3:00 Dynamic schemas and data model churn
- 6:00 Surveillance and privacy
- 13:00 Hadoop, especially the distro wars
- 22:00 BI innovation
- 23:30 More on dynamic schemas and data model churn
Edit: Some of my remarks were transcribed.
- I posted on dynamic schemas data model churn a few days ago.
- I capped off a series on privacy and surveillance a few days ago.
- I commented on various Hadoop distributions in June.
|Categories: Business intelligence, ClearStory Data, Data warehousing, Hadoop, MapR, MapReduce, Surveillance and privacy||Leave a Comment|
I chatted yesterday with the Hortonworks gang. The main subject was Hortonworks’ approach to SQL-on-Hadoop — commonly called Stinger — but at my request we cycled through a bunch of other topics as well. Company-specific notes include:
- Hortonworks founder J. Eric “Eric14″ Baldeschwieler is no longer at Hortonworks, although I imagine he stays closely in touch. What he’s doing next is unspecified, except by the general phrase “his own thing”. (Derrick Harris has more on Eric’s departure.)
- John Kreisa still is at Hortonworks, just not as marketing VP. Think instead of partnerships and projects.
- ~250 employees.
- ~70-75 subscription customers.
Our deployment and use case discussions were a little confused, because a key part of Hortonworks’ strategy is to support and encourage the idea of combining use cases and workloads on a single cluster. But I did hear:
- 10ish nodes for a typical starting cluster.
- 100ish nodes for a typical “data lake” committed adoption.
- Teradata UDA (Unified Data Architecture)* customers sometimes (typically?) jumping straight to a data lake scenario.
- A few users in the 10s of 1000s of nodes. (Obviously Yahoo is one.)
- HBase used in >50% of installations.
- Hive probably even more than that.
- Hortonworks is seeing a fair amount of interest in Windows Hadoop deployments.
*By the way — Teradata seems serious about pushing the UDA as a core message.
Ecosystem notes, in Hortonworks’ perception, included:
- Cloudera is obviously Hortonworks’ biggest distro competitor. Next is IBM, presumably in its blue-forever installed base. MapR is barely on the radar screen; Pivotal’s likely rise hasn’t yet hit sales reports.
- Hortonworks evidently sees a lot of MicroStrategy and Tableau, and some Platfora and Datameer, the latter two at around the same level of interest.
- Accumulo is a big deal in the Federal government, and has gotten a few health care wins as well. Its success is all about security. (Note: That’s all consistent with what I hear elsewhere.)
I also asked specifically about OpenStack. Hortonworks is a member of the OpenStack project, contributes nontrivially to Swift and other subprojects, and sees Rackspace as an important partner. But despite all that, I think strong Hadoop/OpenStack integration is something for the indefinite future.
Hortonworks’ views about Hadoop 2.0 start from the premise that its goal is to support running a multitude of workloads on a single cluster. (See, for example, what I previously posted about Tez and YARN.) Timing notes for Hadoop 2.0 include:
- It’s been in preview/release candidate/commercial beta mode for weeks.
- Q3 is the goal; H2 is the emphatic goal.
- Yahoo’s been in production with YARN >8 months, and has no MapReduce 1 clusters left. (Yahoo has >35,000 Hadoop nodes.)
- The last months of delays have been mainly about sprucing up various APIs and protocols, which may need to serve for a similar multi-year period as Hadoop 1’s have. But there also was some YARN stabilization into May.
Frankly, I think Cloudera’s earlier and necessarily incremental Hadoop 2 rollout was a better choice than Hortonworks’ later big bang, even though the core-mission aspect of Hadoop 2.0 is what was least ready. HDFS (Hadoop Distributed File System) performance, NameNode failover and so on were well worth having, and it’s more than a year between Cloudera starting supporting them and when Hortonworks is offering Hadoop 2.0.
Hortonworks’ approach to doing SQL-on-Hadoop can be summarized simply as “Make Hive into as good an analytic RDBMS as possible, all in open source”. Key elements include: Read more
I visited Cloudera Friday for, among other things, a chat about Impala with Marcel Kornacker and colleagues. Highlights included:
- Impala is meant to someday be a competitive MPP (Massively Parallel Processing) analytic RDBMS.
- At the moment, it is not one. For example, Impala lacks any meaningful form of workload management or query optimization.
- While Impala will run against any HDFS (Hadoop Distributed File System) file format, claims of strong performance assume that the data is in Parquet …
- … which is the replacement for the short-lived Trevni …
- … and which for most practical purposes is true columnar.
- Impala is also meant to be more than an RDBMS; Parquet and presumably in the future Impala can accommodate nested data structures.
- Just as Impala runs against most or all HDFS file formats, Parquet files can be used by most Hadoop execution engines, and of course by Pig and Hive.
- The Impala roadmap includes workload management, query optimization, data skipping, user-defined functions, hash distribution, two turtledoves, and a partridge in a pear tree.
Data gets into Parquet via batch jobs only — one reason it’s important that Impala run against multiple file formats — but background format conversion is another roadmap item. A single table can be split across multiple formats — e.g., the freshest data could be in HBase, with the rest is in Parquet.
A few days ago I posted Daniel Abadi’s thoughts in a discussion of Hadapt, Microsoft PDW (Parallel Data Warehouse)/PolyBase, Pivotal/Greenplum Hawq, and other SQL-Hadoop combinations. This is Dave DeWitt’s response. Emphasis mine.
|Categories: Benchmarks and POCs, Cloudera, Clustering, Data warehousing, Greenplum, Hadapt, Hadoop, MapReduce, Microsoft and SQL*Server, PostgreSQL, SQL/Hadoop integration||6 Comments|
The genesis of this post is:
- Dave DeWitt sent me a paper about Microsoft Polybase.
- I argued with Dave about the differences between Polybase and Hadapt.
- I asked Daniel Abadi for his opinion.
- Dan agreed with Dave, in a long email …
- … that he graciously permitted me to lightly-edit and post.
I love my life.
Per Daniel (emphasis mine): Read more
|Categories: Aster Data, Data warehousing, Greenplum, Hadapt, Hadoop, MapReduce, Microsoft and SQL*Server, SQL/Hadoop integration, Theory and architecture||12 Comments|
My client Syncsort:
- Is an ETL (Extract/Transform/Load) vendor, whose flagship product DMExpress was evidently renamed to DMX.
- Has a strong history in and fondness for sort.
- Has announced a new ETL product, DMX-h ETL Edition, which uses Hadoop MapReduce to parallelize DMX by controlling a copy of DMX that resides on every data node of the Hadoop cluster.*
- Has also announced the closely-related DMX-h Sort Edition, offering acceleration for the sorts inherent in Map and Reduce steps.
- Contributed a patch to Apache Hadoop to open up Hadoop MapReduce to make all this possible.
*Perhaps we should question Syncsort’s previous claims of having strong multi-node parallelism already.
The essence of the Syncsort DMX-h ETL Edition story is:
- DMX-h inherits the various ETL-suite trappings of DMX.
- Syncsort claims DMX-h has major performance advantages vs., for example, Hive- or Pig-based alternatives.
- With a copy of DMX on every node, DMX-h can do parallel load/export.
Hadoop 2.0/YARN is the first big step in evolving Hadoop beyond a strict Map/Reduce paradigm, in that it at least allows for the possibility of non- or beyond-MapReduce processing engines. While YARN didn’t meet its target of general availability around year-end 2012, Arun Murthy of Hortonworks told me recently that:
- Yahoo is a big YARN user.
- There are other — paying — YARN users.
- YARN general availability is now targeted for well before the end of 2013.
Arun further told me about Tez, the next-generation Hadoop processing engine he’s working on, which he also discussed in a recent blog post:
With the emergence of Apache Hadoop YARN as the basis of next generation data-processing architectures, there is a strong need for an application which can execute a complex DAG [Directed Acyclic Graph] of tasks which can then be shared by Apache Pig, Apache Hive, Cascading and others. The constrained DAG expressible in MapReduce (one set of maps followed by one set of reduces) often results in multiple MapReduce jobs which harm latency for short queries (overhead of launching multiple jobs) and throughput for large-scale queries (too much overhead for materializing intermediate job outputs to the filesystem). With Tez, we introduce a more expressive DAG of tasks, within a single application or job, that is better aligned with the required processing task – thus, for e.g., any given SQL query can be expressed as a single job using Tez.
This is similar to the approach of BDAS Spark:
Rather than being restricted to Maps and Reduces, Spark has more numerous primitive operations, including map, reduce, sample, join, and group-by. You can do these more or less in any order.
although Tez won’t match Spark’s richer list of primitive operations.
More specifically, there will be six primitive Tez operations:
- HDFS (Hadoop Distributed File System) input and output.
- Sorting on input and output (I’m not sure why that’s two operations rather than one).
- Shuffling of input and output (ditto).
A Map step would compound HDFS input, output sorting, and output shuffling; a Reduce step compounds — you guessed it! — input sorting, input shuffling, and HDFS output.
I can’t think of much in the way of algorithms that would be logically impossible in MapReduce yet possible in Tez. Rather, the main point of Tez seems to be performance, performance consistency, response-time consistency, and all that good stuff. Specific advantages that Arun and I talked about included:
- The requirement for materializing (onto disk) intermediate results that you don’t want to is gone. (Yay!)
- Hadoop jobs will step on each other’s toes less. Instead of Maps and Reduces from unrelated jobs getting interleaved, all the operations from a single job will by default be executed in one chunk. (Even so, I see no reason to expect early releases of Tez to do a great job on highly concurrent mixed workload management.)
- Added granularity brings opportunities for additional performance enhancements, for example in the area of sorting. (Arun loves sorts.)
|Categories: Databricks, Spark and BDAS, Hadoop, Hortonworks, MapReduce, Workload management, Yahoo||14 Comments|
Spark and Shark are interesting alternatives to MapReduce and Hive. At a high level:
- Rather than persisting data to disk after every step, as MapReduce does, Spark instead writes to something called RDDs (Resilient Distributed Datasets), which can live in memory.
- Rather than being restricted to maps and reduces, Spark has more numerous primitive operations, including map, reduce, sample, join, and group-by. You can do these more or less in any order. All the primitives are parallel with respect to the RDDs.
- Shark is a lot like Hive, only rewritten (in significant parts) and running over Spark.
- There’s an approach to launching tasks quickly — ~5 milliseconds or so — that I unfortunately didn’t grasp.
The key concept here seems to be the RDD. Any one RDD:
- Is a collection of Java objects, which should have the same or similar structure.
- Can be partitioned/distributed and shuffled/redistributed across the cluster.
- Doesn’t have to be entirely in memory at once.
Otherwise, there’s a lot of flexibility; an RDD can be a set of tuples, a collection of XML documents, or whatever other reasonable kind of dataset you want. And I gather that:
- At the moment, RDDs expire at the end of a job.
- This restriction will be lifted in a future release.
|Categories: Data models and architecture, Databricks, Spark and BDAS, Hadoop, MapReduce, Memory-centric data management, Open source, Parallelization, SQL/Hadoop integration||10 Comments|