Google Percolator: MapReduce Demise?

Here is my early thoughts after quickly looking into  Google Percolator and skimming the paper .

Major take-away: massive transactional mutating of tens-petabyte-scale dataset on thousands-node cluster is possible!

MapReduce is still useful for distributed sorts of big-data and few other things, nevertheless it’s “karma” has suffered a blow. Beforehand you could end any MapReduce dispute by saying “well… it works for Google”, however, nowadays before you say it you would hear “well…. it didn’t work for Google”. MapReduce is particularly criticized by having 1) too long latency, 2)too wasteful, requiring full rework of the whole tens-of-PB-scale dataset even if only a fraction of it had been changed and 3) inability to support kinda-real-time data processing (meaning processing documents as they are crawled and updating index appropriately). In short: welcome to disillusionment stage of MapReduce saga. And luckily Hadoop is not only MapReduce, I’m convinced Hadoop will thrive and flourish beyond MapReduce and MapReduce being an important big data tool will be widely used where it really makes sense rather than misused or abused in various ways. Aster Data and remaining MPP startups can relax on the issue a bit.

Probably a topic for another post, but I think MapReduce is best leveraged as ETL tool.

See also http://blog.tonybain.com/tony_bain/2010/09/was-stonebraker-right.html for another view on the issue. There are few others posts already published  on Precolator but I haven’t yet looked into them.

I’m very happy about my SVLC-hypothesis, I think I knew it for a long time, but somehow only now, after I have put it on paper, I felt that the reasoning about different analytics approaches became easier. It is like having a map instead of visualizing it. So where is Percolator in the context of SVLC? If it is still considered analytics, Percolator is an SVC system – giving up latency for everything else, albeit to a lot lesser degree than its successor MapReduce. That said Percolator has a sizable part that is not analytics anymore but rather transaction processing. And transaction processing  is not usefully modeled by my SVLC-hypothesis. In summary: Percolator is essentially the trade-off as MapReduce – sacrificing latency for volume-cost-sophistication but more temperate, more rounded,  less-radical.

Unfortunately I haven’t enough time to enjoy the paper as it should be enjoyed with easy weekend-style reading. So inaccuracies may have been infiltrated in:

  • Percolator is big-data ACID-compliant transaction-processing non-relational DBMS.
  • Percolator fits most NoSQL definitions and therefore it is a NoSQL.
  • Percolator  continuously mutates dataset (called data corpus in the paper) with full transactional semantics and in the sizes of tens of petabytes on thousands of nodes.
  • Percolator uses a message-queue style approach for processing crawled data. Meaning, it processes the crawled pages continuously as they arrive updating the index database transactionaly.
  • BEFORE Percolator: Indexing was done in stages taking weeks. All crawled data was accumulated and staged first, then pass-after-pass transformed into index. 100-passes were quoted in the paper as I remember. When the cycle was completed a new one was initiated. Few weeks latency after content was published and before it appeared in Google search results were considered too long in twitter age, so Google implemented some shortcuts allowing preliminary results to show in search before the cycle is completed.
  • Percolator  doesn’t have declarative query language.
  • No joins.
  • Self-stated ~3% single node efficiency relative to the state-of-the-art DBMS system on single node. That’s the price for handling (which is transactional mutating) high-volume dataset… and relatively cheaply. Kudos for Google to being so open on this and not exercising in term-obfuscation. On the other hand, they can afford it… they don’t have to sell it tomorrow on rather competitive NoSQL market ;)
  • Thread-per-transaction model. Heavily threaded many-core servers as I understand.

Architecturally reminds me MoM (Message Oriented Middleware) with transactional queues and guarantied delivery.

Definitely to be continued…

other Percolator blog posts:

http://www.infoq.com/news/2010/10/google-percolator
http://www.theregister.co.uk/2010/09/24/google_percolator
http://coolthingoftheday.blogspot.com/2010/10/percolator-question-is-how-does-google.html
http://www.quora.com/What-is-Google-Percolator

CAP equivalent for analytics?

CAP theorem deals with trade-off in transactional system. It doesn’t need an introduction, unless of course you have been busy on the moon for last couple of years. In this case you can easily Google for good intros. Here is a wikipedia entry on the subject.

I was thinking how would I build an ideal analytics system. Quickly came realization that all “care abouts” cannot be satisfied simultaneously, even assuming enough time for development. Some desirable properties must be sacrificed in favor for others, hence architectural trade-offs are unavoidable in principle. I immediately had déjà vu regarding CAP. So the following is my take on the subject:

SVLC hypothesis regarding architectural trade-offs in analytics

I haven’t came to rigorous definition yet, here is an intuitive one:  Current technology doesn’t allow implementation of a single analytics system that is SVLC which is simultaneously sophisticated, high-volume, low-latency and low-cost .One of these four properties must be sacrificed, the extent to which it is sacrificied determines the extend to which other properties could potentially be implemented.

Deep dive for the brave souls

Let’s reiterate the desired system properties first (see ideal analytics system):

  1. Deep Sophistication => …free-form SQL:2008 with multi-way joins of 2 and more big tables, sorts of big tables and all the rest of data heavy lifting.
  2. High Volume => …handling big data volumes, Let’s cap it 1PB meanwhile for easier thinking.
  3. Low Latency => …subsecond response time for the query on average. A more concrete description is that latency must be low enough to allow analyst working interactively in conversational manner with the system.
  4. Low-Cost => … I’ll define it as commodity hardware and software must not exceed hardware costs. More rigorously? $1/GB/month for actively queried data is my very rough estimation for low-cost.
  5. Multi-form => any data, relational, serialized objects, text etc….
  6. Security => can speak for itself

I found that multi-formness and security doesn’t interfere with implementing the rest of properties and can in principal always be implemented in satisfactory way without major compromises. Some nuances exists tough, but I’ll ignore them for clearness. So removing them and getting the following list:

1. Sophistication (deep)              => S

2. Volume (high)                          => V

3. Latency (low)                           => L

4. Cost (low)                                  => C

These four are highly inter-related and form а constraint system . Implementing one to full extent hampers the rest. Let’s see what trade-offs we have here. Four properties that is 6 potential simple 2-extremes trade-offs. Let’s settle on geometric tetrahedron to model the architectural trade-off space. Four properties correspond to four vertexes and six trade-offs correspond to six edges. Then we model particular trade-off by putting a point on the corresponding edge. So we get something like this:

Okey, so far so good. Now, I’ll try to be а devil advocate and challenge my point that any trade-offs are necessary in the first place. So let’s review the system denoted as

SVLC=> high-volume, low-latency, deep analytics, low cost

Because it is low-latency it will need I/O throughput adequate to scan whole dataset quickly and since it is high-volume (see above for quantitative definition) meaning aforementioned dataset is big, it will need a large number of individual nodes in cluster to provide the required aggregated adequate I/O throughput. The number of machines is further increased with low-cost requirement meaning that simpler servers that are in mainstream sweetspot must be purchased. Therefore system becomes extremely distributed and data being dispersed all over it. The low-cost networking usually mean TCP/IP that is high-overhead, high-latency and low-throughput. Deep Sophistication analytics requires performing complex data-intensive operations like full sorting of big datasets, joining big tables or just simple select distinct over big data that will inevitably have long latencies. Once latency is long enough that probability of node failing mid-query become non-neglectable. Latency increase becomes self-perpetuating because of required finer grain of  intermediate result materialization. This is needed to prevent never-ending query restarting and provide a kind of resumable queries. Not other solutions to resumable queries are documented except MapReduce-style intermediate result materialization. This ultimately makes latency batch-class long violating low-latency requirement.

I guess my proof miss the required rigor to be considered seriously by academics, I’m just an engineer :) I love to see it reworked to something more serious tough. I just hope to get the point across and to be of value to engineers and practicing architects.

Anyhow this is the base of my hypothesis showing that it impossible to achieve full SVLC using today technology.

Let’s consider other cases where we give up something. It is easy to visualize such trade-off as a 2D plane dissecting tetrahedron. The three points were three edges are cut corresponds to three trade-offs. For simplicity I’ll elaborate only radical trade-offs in this post. Radical trade-offs are those were on all six trade-off edges one extreme is selected and this corresponds to putting a trade-off point on one of vertices. Most real-world system make temperate trade-offs that corresponds to the plane that dissects the tetrahedron into sizable parts. Moreover real-world system, especially available from commercial vendors, are a toolbox kind of a system. Meaning that system consists of a set of tools where each one makes a different set of trade-offs. Then it is up to engineer to choose the right tool for the job to the toolbox. However, toolbox approach is not a loophole for this hypothesis, because properties of the different tools don’t add up in desired way. For example the simultaneous use of expensive tool and low-cost tool is still expensive; the simultaneous use of low-latency and high-latency is high-latency. Nevertheless, toolbox approach is best one for real-world problems. Because real-world problems are usually decomposed to a number of sub problems where each may require different tool.

Well…back to the radical systems… Let’s consider all four cases where we completely give up one property to max out the rest three:


SVL => high-volume, low-latency, deep analytics …giving up Cost… seems to be implementable. In its pure form it reminds classic national security style system. Subsecond querying petabyte-scale dataset with arbitrary joins. Heavily over provisioned Netezza / Exadata / Greenplum / Aster and other MPP-system could do it I believe. Data kept in huge RAM or on flash, huge I/O is available to scan the whole dataset in matter of seconds. High-speed, low-overhead networking is available to with huge bi-section network bandwidth capable to shuffle the whole dataset in matter of seconds. Infiniband/RDMA are the best probably. How bad Cost can be here? Well… unhealthy to imagine. Throw some numbers in anyway? Will do some back of envelope calculation in my future posts.

SVC => high-volume, low-cost, deep analytics …giving up Latency… seems to be implementable, in fact it is MapReduce territory, Hadoop natural habitat. Are ETL systems SVC? I think no, because while they given up Latency they haven’t kept on Volume. How bad is Latency? well… forget interactivity, create queuing system and get notified when the job is done. If too slow add servers. If some interactive experimentation is needed, use VLC first to develop and prove your hypothesis and only than crunch the data with SVC. Since cost is involved I guess Hadoop MapReduce is really a king here. Tough if Aster licenses for example are comparable to commodity cluster overall cost and is not many multiples of it, then it could fit the category nicely. Otherwise it will make suboptimal (considering my model context not in wider sense!) great SV system. The great MapReduce debate is not for nothing!

SLC => low-latency, low-cost, deep analytics …seems to be implementable in a minute, just start your favorite spreadsheet application ;) You will be shocked how much data Excel crunches in just few seconds, nowadays. Most traditional BI tools are in this category too. Heck, if not for BigData, the analytics industry will be as would become as exciting as enterprise payroll systems. Though, innovation is possible even there.  Heck, 99% of BI is fully feasible to be done completely  in-memory, often on single server and the deployment must be really low-risk low-cost very-rapid if done correctly. Most cloud BI vendors are also in this category. “R-project” is here too. This was Kickfire beloved spot as well as is now for QlikTech & GoofData, PivotLink and etc… So pretty much all BI vendors are here except MPP heavy-lifters. How bad Volumes are limited? Well with CPU-DRAM bandwidth being 50GB/sec and DRAM sizes 64GB on common commodity servers I think crunching few tens GB should be well possible in matter of seconds, if not for implementation sloppiness, and with literally pocket money (average enterprise’s pocket not mine… yet).

VLC => high-volume, low-latency, low-cost …giving up Sophistication…seems to be implementable, that is doing a simple scan and giving up the Sophistication, particularly joins. Dremel and BigQuery seem following this approach. How bad is giving up Sophistication? Well, it all depends on how pre-joined/nested dataset is. With normalized schemas, well… unavailability of joins makes it pretty much impractical implement any usable analytics. However, with star-schema and particularly nested data (with some extensive pre-joins even if it means some redundancy), this can work wonders to vast majority of queries, completing them in seconds on even large datasets. However, no pre-join strategy will work for 100% of queries and functions like COUNT DISTINCT must be approximated when run over big dataset like described in Dremel paper. Also I’ll assign sampling strategy to this category, because sophistication also means accuracy here. One clarification: only joins of several big tables are sacrificed here, joins of big table with even large number of small tables are perfectly okay and done on the fly during the scan. Sorts of big table before it was reduced significantly to manageable size is also sacrificed in this approach, however approximation algorithms can be used for this and then it will be okay too.

Hence the conclusion: only 3 of 4 SVLC properties can be implemented in full extend in single analytic system. The hypothesis goes that any attempt that allegedly violates it, in fact either is no a single system or impairment is latent in one or more properties.

[TODO: rewrite] The extended hypothesis for fractional cases:

  • Systems/trade-offs may be radical or temperate. Radical trade-off completely gives up one of four properties of the system. Temperate trade-off gives-up the property only fractionally on expense of giving-up other properties also fractionally.
  • Most real-world systems are complex. They are a set of tools, where each separate tool is a concrete trade-off. Then the user of such system can use different tools with different trade-off sequentially or simultaneously. This may seems as way out of the restraint; however it is not, because properties of separate tools don’t add up in desired way. For example the simultaneous use of expensive tool and low-cost tool is still expensive; the simultaneous use of low-latency and high-latency is high-latency.  Nevertheless, toolbox approach is best one for real-world problems. Because real-world problems are usually decomposed to a number of sub problems where each may require different tool.
  • Most often trade-offs of real-world systems are temperate.