The pessimist thinks the glass is half-empty.
The optimist thinks the glass is half-full.
The engineer thinks the glass was poorly designed.
Most of what I wrote in Part 1 of this post was already true 15 years ago. But much gets added in the modern era, considering that:
- Clusters will have node hiccups more often than single nodes will. (Duh.)
- Networks are relatively slow even when uncongested, and furthermore congest unpredictably.
- In many applications, it’s OK to sacrifice even basic-seeming database functionality.
And so there’s been innovation in numerous cluster-related subjects, two of which are:
- Distributed query and update. When a database is distributed among many modes, how does a request access multiple nodes at once?
- Fault-tolerance in long-running jobs.When a job is expected to run on many nodes for a long time, how can it deal with failures or slowdowns, other than through the distressing alternatives:
- Start over from the beginning?
- Keep (a lot of) the whole cluster’s resources tied up, waiting for things to be set right?
Distributed database consistency
When a distributed database lives up to the same consistency standards as a single-node one, distributed query is straightforward. Performance may be an issue, however, which is why we have seen a lot of:
- Analytic RDBMS innovation.
- Short-request applications designed to avoid distributed joins.
- Short-request clustered RDBMS that don’t allow fully-general distributed joins in the first place.
But in workloads with low-latency writes, living up to those standards is hard. The 1980s approach to distributed writing was two-phase commit (2PC), which may be summarized as: Read more
|Categories: Clustering, CouchDB, Data warehousing, Databricks, Spark and BDAS, Facebook, Hadoop, MapReduce, Sybase, Theory and architecture, VoltDB and H-Store||1 Comment|
Writing data management or analysis software is hard. This post and its sequel are about some of the reasons why.
When systems work as intended, writing and reading data is easy. Much of what’s hard about data management is dealing with the possibility — really the inevitability — of failure. So it might be interesting to survey some of the many ways that considerations of failure come into play. Some have been major parts of IT for decades; others, if not new, are at least newly popular in this cluster-oriented, RAM-crazy era. In this post I’ll focus on topics that apply to single-node systems; in the sequel I’ll emphasize topics that are clustering-specific.
Major areas of failure-aware design — and these overlap greatly — include:
- Backup and restore. In its simplest form, this is very basic stuff. That said — any decent database management system should let backups be made without blocking ongoing database operation, with the least performance impact possible.
- Logging, rollback and replay. Logs are essential to DBMS. And since they’re both ubiquitous and high-performance, logs are being used in ever more ways.
- Locking, latching, transactions and consistency. Database consistency used to be enforced in stern and pessimistic ways. That’s changing, big-time, in large part because of the requirements of …
- … distributed database operations. Increasingly, modern distributed database systems are taking the approach of getting work done first, then cleaning up messes when they occur.
- Redundancy and replication. Parallel computing creates both a need and an opportunity to maintain multiple replicas of data at once, in very different ways than the redundancy and replication of the past.
- Fault-tolerant execution. When one node is inoperative, inaccessible, overloaded or just slow, you may not want a whole long multi-node job to start over. A variety of techniques address this need.
In a single-server, disk-based configuration, techniques for database fault-tolerance start: Read more
As per the links and quotes below, my views on the network neutrality debate may be summarized as:
- There should be a fairly good level of internet delivery that is, by regulation, available to any website or other internet service. This is essential so that ideas can blossom, speech can be shared, etc.
- There should be a way to pay for arbitrarily good levels of internet delivery. Entertainment would benefit from that. Medicine, in the future, might require it.
- Any payment for better delivery should happen through a marketplace open to all.
In this post I’ll add detail as to how that marketplace could work.
After visiting California recently, I made a flurry of posts, several of which generated considerable discussion.
- My claim that Spark will replace Hadoop MapReduce got much Twitter attention — including some high-profile endorsements — and also some responses here.
- My MemSQL post led to a vigorous comparison of MemSQL vs. VoltDB.
- My post on hardware and storage spawned a lively discussion of Hadoop hardware pricing; even Cloudera wound up disagreeing with what I reported Cloudera as having said. Sadly, there was less response to the part about the partial (!) end of Moore’s Law.
- My Cloudera/SQL/Impala/Hive apparently was well-balanced, in that it got attacked from multiple sides via Twitter & email. Apparently, I was too hard on Impala, I was too hard on Hive, and I was too hard on boxes full of cardboard file cards as well.
- My post on the Intel/Cloudera deal garnered a comment reminding us Dell had pushed the Intel distro.
- My CitusDB post picked up a few clarifying comments.
Here is a catch-all post to complete the set. Read more
One of my lesser-known clients is Citus Data, a largely Turkish company that is however headquartered in San Francisco. They make CitusDB, which puts a scale-out layer over a collection of fully-functional PostgreSQL nodes, much like Greenplum and Aster Data before it. However, in contrast to those and other Postgres-based analytic MPP (Massively Parallel Processing) DBMS:
- CitusDB does not permanently fork PostgreSQL; Citus Data has committed to always working with the latest PostgreSQL release, or at least with one that’s less than a year old.
- Citus Data never made the “fat head” mistake — if a join can’t be executed directly on the CitusDB data-storing nodes, it can’t be executed in CitusDB at all.
- CitusDB follows the modern best-practice of having many virtual nodes on each physical node. Default size of a virtual node is one gigabyte. Each virtual node is technically its own PostgreSQL table.*
- Citus Data has already introduced an open source column-store option for PostgreSQL, which CitusDB of course exploits.
*One benefit to this strategy, besides the usual elasticity and recovery stuff, is that while PostgreSQL may be single-core for any given query, a CitusDB query can use multiple cores by virtue of hitting multiple PostgreSQL tables on each node.
Citus has thrown a few things against the wall; for example, there are two versions of its product, one which involves HDFS (Hadoop Distributed File System) and one of which doesn’t. But I think Citus’ focus will be scale-out PostgreSQL for at least the medium-term future. Citus does have actual customers, and they weren’t all PostgreSQL users previously. Still, the main hope — at least until the product is more built-out — is that existing PostgreSQL users will find CitusDB easy to adopt, in technology and price alike.
|Categories: Aster Data, Citus Data, Columnar database management, Data warehousing, Database compression, Greenplum, Hadoop, Parallelization, PostgreSQL, SQL/Hadoop integration, Transparent sharding, Workload management||6 Comments|
I stopped by MemSQL last week, and got a range of new or clarified information. For starters:
- Even though MemSQL (the product) was originally designed for OLTP (OnLine Transaction Processing), MemSQL (the company) is now focused on analytic use cases …
- … which was the point of introducing MemSQL’s flash-based columnar option.
- One MemSQL customer has a 100 TB “data warehouse” installation on Amazon.
- Another has “dozens” of terabytes of data spread across 500 machines, which aggregate 36 TB of RAM.
- At customer Shutterstock, 1000s of non-MemSQL nodes are monitored by 4 MemSQL machines.
- A couple of MemSQL’s top references are also Vertica flagship customers; one of course is Zynga.
- MemSQL reports encountering Clustrix and VoltDB in a few competitive situations, but not NuoDB. MemSQL believes that VoltDB is still hampered by its traditional issues — Java, reliance on stored procedures, etc.
On the more technical side: Read more
|Categories: Clustering, Clustrix, Columnar database management, Data warehousing, Database compression, In-memory DBMS, MemSQL, NewSQL, NuoDB, Specific users, Vertica Systems, VoltDB and H-Store, Workload management, Zynga||16 Comments|
My California trip last week focused mainly on software — duh! — but I had some interesting hardware/storage/architecture discussions as well, especially in the areas of:
- Rack- or data-center-scale systems.
- The real or imagined demise of Moore’s Law.
I also got updated as to typical Hadoop hardware.
If systems are designed at the whole-rack level or higher, then there can be much more flexibility and efficiency in terms of mixing and connecting CPU, RAM and storage. The Google/Facebook/Amazon cool kids are widely understood to be following this approach, so others are naturally considering it as well. My most interesting of several mentions of that point was when I got the chance to talk with Berkeley computer architecture guru Dave Patterson, who’s working on plans for 100-petabyte/terabit-networking kinds of systems, for usage after 2020 or so. (If you’re interested, you might want to contact him; I’m sure he’d love more commercial sponsorship.)
One of Dave’s design assumptions is that Moore’s Law really will end soon (or at least greatly slow down), if by Moore’s Law you mean that every 18 months or so one can get twice as many transistors onto a chip of the same area and cost than one could before. However, while he thinks that applies to CPU and RAM, Dave thinks flash is an exception. I gathered that he thinks the power/heat reasons for Moore’s Law to end will be much harder to defeat than the other ones; note that flash, because of what it’s used for, has vastly less power running through it than CPU or RAM do.
|Categories: Amazon and its cloud, Buying processes, Cloudera, Facebook, Google, Intel, Memory-centric data management, Pricing, Solid-state memory||17 Comments|
Intel recently made a huge investment in Cloudera, stated facts about which start:
- $740 million …
- … for 18% of the company …
- … as part of an overall $900 million round.
- SEC filings coming soon with more details.
Give or take stock preferences, etc., that’s around a $4.1 billion valuation post-money, but Cloudera does say it now has “most of $1 billion” in the bank.
Cloudera further told me when I visited last Friday that the majority of the Intel investment is net new money. (I presume that the rest of the round is net-new as well.) Hence, I conclude that previous investors sold in the aggregate less than 10% of total holdings to Intel. While I’m pretty sure Mike Olson is buying himself a couple of nice toys, in most respects it’s business-as-usual at Cloudera, with the same investors, directors and managers they had before. By way of contrast, many of the “cashing-out” rumors going around are OBVIOUSLY absurd, unless you think Intel acquired a much larger fraction of Cloudera than it actually did.
That said, Intel spent a lot of money, and in connection with the investment there’s a tight Cloudera/Intel partnership. In particular, Read more
There’s much confusion about Cloudera’s SQL plans and beliefs, and the company has mainly itself to blame. That said, here’s what I think is going on.
- Hive is good at some tasks and terrible at others.
- Hive is good at batch data transformation.
- Hive is bad at ad-hoc query, unless you really, really need Hive’s scale and low license cost. One example, per Eli Collins: Facebook has a 500 petabyte Hive warehouse, but jokes that on a good day an analyst can run 6 queries against it.
- Impala is meant to be good at what Hive is bad at – i.e., fast-response query. (Cloudera mentioned reliable 100 millisecond response times for at least one user.)
- Impala is also meant to be good at what Hive is good at, and will someday from Cloudera’s standpoint completely supersede Hive, but Cloudera is in no hurry for that day to arrive. Hive is more mature. Hive still has more SQL coverage than Impala. There’s a lot of legacy investment in Hive. Cloudera gets little business advantage if a customer sunsets Hive.
- Impala is already decent at some tasks analytic RDBMS are commonly used for. Cloudera insists that some queries run very quickly on Impala. I believe them.
- Impala is terrible at others, including some of the ones most closely associated with the concept of “data warehousing”. Data modeling is a big zero right now. Impala’s workload management, concurrency and all that are very immature.
- There are some use cases for which SQL-on-Hadoop blows away analytic RDBMS, for example ones involving data transformations – perhaps on multi-structured data – that are impractical in RDBMS.
And of course, as vendors so often do, Cloudera generally overrates both the relative maturity of Impala and the relative importance of the use cases in which its offerings – Impala or otherwise – shine.
- A survey of SQL/Hadoop integration (February, 2014)
- The cardinal rules of DBMS development (March, 2013)
|Categories: Cloudera, Data warehousing, Facebook, Hadoop, SQL/Hadoop integration, Workload management||4 Comments|
Spark is on the rise, to an even greater degree than I thought last month.
- Numerous clients and other companies I talk with have adopted Spark, plan to adopt Spark, or at least think it’s likely they will. In particular:
- A number of analytic-stack companies are joining ClearStory in using Spark. Most of the specifics are confidential, but I hope some will be announced soon.
- MapR has joined Cloudera in supporting Spark, and indeed — unlike Cloudera — is supporting the full Spark stack.
- Mike Olson of Cloudera is on record as predicting that Spark will be the replacement for Hadoop MapReduce. Just about everybody seems to agree, except perhaps for Hortonworks folks betting on the more limited and less mature Tez. Spark’s biggest technical advantages as a general data processing engine are probably:
- The Directed Acyclic Graph processing model. (Any serious MapReduce-replacement contender will probably echo that aspect.)
- A rich set of programming primitives in connection with that model.
- Support also for highly-iterative processing, of the kind found in machine learning.
- Flexible in-memory data structures, namely the RDDs (Resilient Distributed Datasets).
- A clever approach to fault-tolerance.
- Spark is a major contender in streaming.
- There’s some cool machine-learning innovation using Spark.
- Spark 1.0 will drop by mid-May, Apache voters willin’ an’ the creek don’ rise. Publicity will likely ensue, with strong evidence of industry support.*
*Yes, my fingerprints are showing again.
The most official description of what Spark now contains is probably the “Spark ecosystem” diagram from Databricks. However, at the time of this writing it is slightly out of date, as per some email from Databricks CEO Ion Stoica (quoted with permission):
… but if I were to redraw it, SparkSQL will replace Shark, and Shark will eventually become a thin layer above SparkSQL and below BlinkDB.
With this change, all the modules on top of Spark (i.e., SparkStreaming, SparkSQL, GraphX, and MLlib) are part of the Spark distribution. You can think of these modules as libraries that come with Spark.
|Categories: Cloudera, Complex event processing (CEP), Databricks, Spark and BDAS, Hadoop, Hortonworks, MapR, MapReduce, Predictive modeling and advanced analytics, SQL/Hadoop integration, Yahoo||14 Comments|