A significant fraction of IT professional services industry revenue comes from data integration. But as a software business, data integration has been more problematic. Informatica, the largest independent data integration software vendor, does $1 billion in revenue. INFA’s enterprise value (market capitalization after adjusting for cash and debt) is $3 billion, which puts it way short of other category leaders such as VMware, and even sits behind Tableau.* When I talk with data integration startups, I ask questions such as “What fraction of Informatica’s revenue are you shooting for?” and, as a follow-up, “Why would that be grounds for excitement?”
*If you believe that Splunk is a data integration company, that changes these observations only a little.
On the other hand, several successful software categories have, at particular points in their history, been focused on data integration. One of the major benefits of 1990s business intelligence was “Combines data from multiple sources on the same screen” and, in some cases, even “Joins data from multiple sources in a single view”. The last few years before application servers were commoditized, data integration was one of their chief benefits. Data warehousing and Hadoop both of course have a “collect all your data in one place” part to their stories — which I call data mustering — and Hadoop is a data transformation tool as well.
And it’s not as if successful data integration companies have no value. IBM bought a few EAI (Enterprise Application Integration) companies, plus top Informatica competitor Ascential, plus Cast Iron Systems. DataDirect (I mean the ODBC/JDBC guys, not the storage ones) has been a decent little business through various name changes and ownerships (independent under a couple of names, then Intersolv/Merant, then independent again, then Progress Software). Master data management (MDM) and data cleaning have had some passable exits. Talend raised $40 million last December, which is a nice accomplishment if you’re French.
I can explain much of this in seven words: Data integration is both important and fragmented. The “important” part is self-evident; I gave examples of “fragmented” a couple years back. Beyond that, I’d say:
- A new class of “engine” can be a nice business — consider for example Informatica/Ascential/Ab Initio, or the MDM players (who sold out to bigger ETL companies), or Splunk. Indeed, much early Hadoop adoption was for its capabilities as a data transformation engine.
- Data transformation is a better business to enter than data movement. Differentiated value in data movement comes in areas such as performance, reliability and maturity, where established players have major advantages. But differentiated value in data transformation can come from “intelligence”, which is easier to excel in as a start-up.
- “Transparent connectivity” is a tough business. It is hard to offer true transparency, with minimal performance overhead, among enough different systems for anybody to much care. And without that you’re probably offering a low-value/niche capability. Migration aids are not an exception; the value in those is captured by the vendor of what’s being migrated to, not by the vendor who actually does the transparent translation. Indeed …
- … I can’t think of a single case in which migration support was a big software business. (Services are a whole other story.) Perhaps Cast Iron Systems came closest, but I’m not sure I’d categorize it as either “migration support” or “big”.
And I’ll stop there, because I’m not as conversant with some of the new “smart data transformation” companies as I’d like to be.
- DBMS transparency layers never seem to sell well (April, 2009)
- ClearStory’s approach to data integration (September, 2013)
- Judging opportunities (July, 2014)
Oracle is announcing today what it’s calling “Oracle Big Data SQL”. As usual, I haven’t been briefed, but highlights seem to include:
- Oracle Big Data SQL is basically data federation using the External Tables capability of the Oracle DBMS.
- Unlike independent products — e.g. Cirro — Oracle Big Data SQL federates SQL queries only across Oracle offerings, such as the Oracle DBMS, the Oracle NoSQL offering, or Oracle’s Cloudera-based Hadoop appliance.
- Also unlike independent products, Oracle Big Data SQL is claimed to be compatible with Oracle’s usual security model and SQL dialect.
- At least when it talks to Hadoop, Oracle Big Data SQL exploits predicate pushdown to reduce network traffic.
And by the way – Oracle Big Data SQL is NOT “SQL-on-Hadoop” as that term is commonly construed, unless the complete Oracle DBMS is running on every node of a Hadoop cluster.
Predicate pushdown is actually a simple concept:
- If you issue a query in one place to run against a lot of data that’s in another place, you could spawn a lot of network traffic, which could be slow and costly. However …
- … if you can “push down” parts of the query to where the data is stored, and thus filter out most of the data, then you can greatly reduce network traffic.
“Predicate pushdown” gets its name from the fact that portions of SQL statements, specifically ones that filter data, are properly referred to as predicates. They earn that name because predicates in mathematical logic and clauses in SQL are the same kind of thing — statements that, upon evaluation, can be TRUE or FALSE for different values of variables or data.
The most famous example of predicate pushdown is Oracle Exadata, with the story there being:
- Oracle’s shared-everything architecture created a huge I/O bottleneck when querying large amounts of data, making Oracle inappropriate for very large data warehouses.
- Oracle Exadata added a second tier of servers each tied to a subset of the overall storage; certain predicates are pushed down to that tier.
- The I/O between Exadata’s two sets of servers is now tolerable, and so Oracle is now often competitive in the high-end data warehousing market,
Oracle evidently calls this “SmartScan”, and says Oracle Big Data SQL does something similar with predicate pushdown into Hadoop.
Oracle also hints at using predicate pushdown to do non-tabular operations on the non-relational systems, rather than shoehorning operations on multi-structured data into the Oracle DBMS, but my details on that are sparse.
- Chris Kanaracus’ coverage of the announcement quotes me at length.
As part of my series on the keys to and likelihood of success, I outlined some examples from the DBMS industry. The list turned out too long for a single post, so I split it up by millennia. The part on 20th Century DBMS success and failure went up Friday; in this one I’ll cover more recent events, organized in line with the original overview post. Categories addressed will include analytic RDBMS (including data warehouse appliances), NoSQL/non-SQL short-request DBMS, MySQL, PostgreSQL, NewSQL and Hadoop.
DBMS rarely have trouble with the criterion “Is there an identifiable buying process?” If an enterprise is doing application development projects, a DBMS is generally chosen for each one. And so the organization will generally have a process in place for buying DBMS, or accepting them for free. Central IT, departments, and — at least in the case of free open source stuff — developers all commonly have the capacity for DBMS acquisition.
In particular, at many enterprises either departments have the ability to buy their own analytic technology, or else IT will willingly buy and administer things for a single department. This dynamic fueled much of the early rise of analytic RDBMS.
Buyer inertia is a greater concern.
- A significant minority of enterprises are highly committed to their enterprise DBMS standards.
- Another significant minority aren’t quite as committed, but set pretty high bars for new DBMS products to cross nonetheless.
- FUD (Fear, Uncertainty and Doubt) about new DBMS is often justifiable, about stability and consistent performance alike.
A particularly complex version of this dynamic has played out in the market for analytic RDBMS/appliances.
- First the newer products (from Netezza onwards) were sold to organizations who knew they wanted great performance or price/performance.
- Then it became more about selling “business value” to organizations who needed more convincing about the benefits of great price/performance.
- Then the behemoth vendors became more competitive, as Teradata introduced lower-price models, Oracle introduced Exadata, Sybase got more aggressive with Sybase IQ, IBM bought Netezza, EMC bought Greenplum, HP bought Vertica and so on. It is now hard for a non-behemoth analytic RDBMS vendor to make headway at large enterprise accounts.
- Meanwhile, Hadoop has emerged as serious competitor for at least some analytic data management, especially but not only at internet companies.
Otherwise I’d say:
- At large enterprises, their internet operations perhaps excepted:
- Short-request/general-purpose SQL alternatives to the behemoths — e.g. MySQL, PostgreSQL, NewSQL — have had tremendous difficulty getting established. The last big success was the rise of Microsoft SQL Server in the 1990s. That’s why I haven’t mentioned the term mid-range DBMS in years.
- NoSQL/non-SQL has penetrated large enterprises mainly for a few specific use cases, for example the lists I posted for MongoDB or graph databases.
- Internet-only companies have few inertia issues when it comes to database managers. They’ll consider anything they regard as being in their price ballpark (which is however often restricted to open source). I think part of the reason is that as quickly as they rewrite their applications, DBMS are vastly less “strategic” to them than they are to most larger enterprises.
- The internet operations of large companies — especially large retailers — in many cases behave like internet-only companies, but in many other cases behave like the rest of the enterprise.
The major reasons for DBMS categories to get established in the first place are:
- Performance and/or scalability (many examples).
- Developer features (for example dynamic schema).
- License/maintenance cost (for example several open source categories).
- Ease of installation and administration (for example open source again, and also data warehouse appliances).
Those same characteristics are major bases for competition among members of a new category, although as noted above behemoth-loyalty can also come into play.
Cool-vs.-weird tradeoffs are somewhat secondary among SQL DBMS.
- There’s not much of a “cool” factor, because new products aren’t that different in what they do vs. older ones.
- There’s not a terrible “weird” factor either, but of course any smaller offering faces FUD, and also …
- … appliances are anti-strategic for many buyers, especially ones who demand a smooth path to the cloud.)
They’re huge, however, in the non-SQL world. Most non-SQL data managers have a major “weird” factor. Fortunately, NoSQL and Hadoop both have huge “cool” cred to offset it. XML/XQuery unfortunately did not.
Finally, in most DBMS categories there are massive issues with product completeness, more in the area of maturity than that of whole product. The biggest whole product issues are concentrated on the matter of interoperating with other software — business intelligence tools, packaged applications (if relevant to the category), etc. Most notably, the handful of DBMS that are certified to run SAP share a huge market that other DBMS can’t touch. But BI tools are less of a differentiator — I yawn when vendors tell me they are certified for/partnered with MicroStrategy, Tableau, Pentaho and Jaspersoft, and I’m surprised at any product that isn’t.
DBMS maturity has a lot of aspects, but the toughest challenges are concentrated in two main areas:
- Reliability, especially but not only in short-request use cases.
- Performance across a great variety of use cases. I observe frequently that performance in best-case scenarios, performance in the lab and performance in real-world environments are much further apart than vendors like to think.
- Maturity demands seem to be much higher for SQL DBMS than for NoSQL.
- I think this is one of several reasons NoSQL has been much more successful than NewSQL.
- It’s why I think MarkLogic’s “Enterprise NoSQL” positioning is a mistake.
- As for MySQL:
- MySQL wasn’t close to reliable enough for enterprises to trust it until InnoDB became the default storage engine.
- MySQL 5 point releases have added major features, or decent performance for major features. I’ll confess to having lost track of what’s been fixed and what’s still missing.
- In saying all that I’m holding MySQL to a much higher maturity standard than I’m holding NoSQL — because that’s what I think enterprise customers do.
- PostgreSQL “should” be doing a lot better than it is. I have an extremely low opinion of its promoters, and not just for personal reasons. (That said, the personal reasons don’t just apply to EnterpriseDB anymore. I’ve also run out of patience waiting for Josh Berkus to retract untruths he posted about me years ago.)
- SAP HANA checks boxes for performance (In-memory rah rah rah!!) and whole product (Runs SAP!!). That puts it well ahead of most other newish SQL DBMS, purely analytic ones perhaps excepted.
- Any other new short-request SQL DBMS that sounds like is has traction is also memory-centric.
- Analytic RDBMS are in most respects held to lower maturity standards than DBMS used for write-intensive workloads. Even so, products in the category are still frequently tripped up by considerations of concurrent performance and mixed workload management.
There have been 1,470 previous posts in the 9-year history of this blog, many of which could serve as background material for this one. A couple that seem particularly germane and didn’t get already get linked above are:
I’m commonly asked to assess vendor claims of the kind:
- “Our system lets you do multiple kinds of processing against one database.”
- “Otherwise you’d need two or more data managers to get the job done, which would be a catastrophe of unthinkable proportion.”
So I thought it might be useful to quickly review some of the many ways organizations put multiple data stores to work. As usual, my bottom line is:
- The most extreme vendor marketing claims are false.
- There are many different choices that make sense in at least some use cases each.
Horses for courses
It’s now widely accepted that different data managers are better for different use cases, based on distinctions such as:
- Short-request vs. analytic.
- SQL vs. non-SQL (NoSQL or otherwise).
- Expensive/heavy-duty vs. cheap/easy-to-support.
Vendors are part of this consensus; already in 2005 I observed
For all practical purposes, there are no DBMS vendors left advocating single-server strategies.
Vendor agreement has become even stronger in the interim, as evidenced by Oracle/MySQL, IBM/Netezza, Oracle’s NoSQL dabblings, and various companies’ Hadoop offerings.
Multiple data stores for a single application
We commonly think of one data manager managing one or more databases, each in support of one or more applications. But the other way around works too; it’s normal for a single application to invoke multiple data stores. Indeed, all but the strictest relational bigots would likely agree:
- It’s common and sensible to manage authentication and authorization data in its own data store. Commonly, the data format is LDAP (Lightweight Directory Access Protocol).
- It’s common and sensible to manage the “content” and “e-commerce transaction records” aspects of websites separately.
- Even beyond that case, there are often performance reasons to manage BLOBs (Binary Large OBjects) outside your relational database.
- Internet “interaction” data is also often best managed outside an RDBMS, in part because of its very non-tabular data structures.
The spectacular 2010 JP Morgan Chase outage was largely caused, I believe, by disregard of these precepts.
There also are cases in which applications dutifully get all their data via SQL queries, but send those queries to two or more DBMS. Teradata is proud that its systems can support rather transactional queries (for example in call-center use cases), but the same application may read from and write to a true OTLP database as well.
Further, many OLTP (OnLine Transaction Processing) applications do some fraction of their work via inbound or outbound messaging. Many buzzwords can come into play here, including but not limited to:
- SOA (Service-Oriented Architecture). This is the most current and flexible one.
- EAI (Enterprise Application Integration). This was a hot concept in the late 1990s, but was generally implemented with difficulties that SOA was later designed to alleviate.
- Message-oriented middleware (MOM) and Publish/Subscribe. These are even older, and overlap greatly.
Finally, every dashboard that combines information from different data stores could be assigned to this category as well.
Multiple storage approaches in a single DBMS
In theory, a single DBMS could operate like two or more different ones glued together. A few functions should or must be centralized, such as administration, and communication with the outside world (connection handling, parsing, etc.). But data storage, query execution and so on could for the most part be performed by rather loosely coupled subsystems. And so you might have the best of both worlds — something that’s multiple data stores in the ways you want that diversity, but a single system in how it fits into your environment.
I discussed this idea last year with cautious optimism, writing:
So will these trends succeed? The forgoing caveats notwithstanding, my answers are more Yes than No.
- … multi-purpose DBMS will likely always have performance penalties, but over time the penalties should become small enough to be affordable in most cases.
- Machine-generated data and “content” both call for multi-datatype DBMS. And taken together, those are a large fraction of the future of computing. Consequently …
- … strong support for multiple datatypes and DMLs is a must for “general-purpose” RDBMS. Oracle and IBM [have] been working on that for 20 years already, with mixed success. I doubt they’ll get much further without a thorough rewrite, but rewrites happen; one of these decades they’re apt to get it right.
In 2005 I had been more ambivalent, in part because my model was a full 1990s-dream “universal” DBMS:
IBM, Oracle, and Microsoft have all worked out ways to have integrated query parsing and query optimization, while letting storage be more or less separate. More precisely, Oracle actually still sticks everything into one data store (hence the lack of native XML support), but allows near-infinite flexibility in how it is accessed. Microsoft has already had separate servers for tabular data, text, and MOLAP, although like Sybase, it doesn’t have general datatype extensibility that it can expose to customers, or exploit itself to provide a great variety of datatypes. IBM has had Oracle-like extensibility all along, although it hasn’t been quite as aggressive at exploiting it; now it’s introduced a separate-server option for XML.
That covers most of the waterfront, but I’d like to more explicitly acknowledge three trends:
- Among other things, Hadoop is a collection of DBMS (HBase, Impala, et al.) that in some cases are very loosely coupled to each other. The question is less how well the various data stores work together, and more how mature any one of them is on its own.
- The multiple-data-models idea has been extended into schema-on-need, which is sometimes but not always housed in Hadoop.
- Even on the relational side, multiple storage capabilities exist in one product.
- Vertica was designed that way from the get-go. (Like the old joke about police duos, one is to read and one is to write.)
- IBM, Microsoft and Oracle have all recently added some kind of in-memory columnar capability.
- Teradata, Aster (before Teradata bought them), Greenplum and Vertica all added some variant on row/column dual stores.
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:
- A write is planned and parceled out to occur on all the different nodes where the data needs to be placed.
- Each node decides it’s ready to commit the write.
- Each node informs the others of its readiness.
- Each node actually commits.
Unfortunately, if any of the various messages in the 2PC process is delayed, so is the write. This creates way too much likelihood of work being blocked. And so modern approaches to distributed data writing are more … well, if I may repurpose the famous Facebook slogan, they tend to be along the lines of “Move fast and break things”,* with varying tradeoffs among consistency, other accuracy, reliability, functionality, manageability, and performance.
By the way — Facebook recently renounced that motto, in favor of “Move fast with stable infrastructure.” Hmm …
Back in 2010, I wrote about various approaches to consistency, with the punch line being:
A conventional relational DBMS will almost always feature RYW consistency. Some NoSQL systems feature tunable consistency, in which — depending on your settings — RYW consistency may or may not be assured.
The core ideas of RYW consistency, as implemented in various NoSQL systems, are:
- Let N = the number of copies of each record distributed across nodes of a parallel system.
- Let W = the number of nodes that must successfully acknowledge a write for it to be successfully committed. By definition, W <= N.
- Let R = the number of nodes that must send back the same value of a unit of data for it to be accepted as read by the system. By definition, R <= N.
- The greater N-R and N-W are, the more node or network failures you can typically tolerate without blocking work.
- As long as R + W > N, you are assured of RYW consistency.
That bolded part is the key point, and I suggest that you stop and convince yourself of it before reading further.
Eventually , Dan Abadi claimed that the key distinction is synchronous/asynchronous — is anything blocked while waiting for acknowledgements? From many people, that would simply be an argument for optimistic locking, in which all writes go through, and conflicts — of the sort that locks are designed to prevent — cause them to be rolled back after-the-fact. But Dan isn’t most people, so I’m not sure — especially since the first time I met Dan was to discuss VoltDB predecessor H-Store, which favors application designs that avoid distributed transactions in the first place.
One idea that’s recently gained popularity is a kind of semi-synchronicity. Writes are acknowledged as soon as they arrive at a remote node (that’s the synchronous part). Each node then updates local permanent storage on its own, with no further confirmation. I first heard about this in the context of replication, and generally it seems designed for replication-oriented scenarios.
Finally, let’s consider fault-tolerance within a single long-running job, whether that’s a big query or some other kind of analytic task. In most systems, if there’s a failure partway through a job, they just say “Oops!” and start it over again. And in non-extreme cases, that strategy is often good enough.
Still, there are a lot of extreme workloads these days, so it’s nice to absorb a partial failure without entirely starting over.
- Hadoop MapReduce, which stores intermediate results anyway, finds it easy to replay just the parts of the job that went awry.
- Spark, which is more flexible in execution graph and data structures alike, has a similar capability.
Additionally, both Hadoop and Spark support speculative execution, in which several clones of a processing step are executed at once (presumably on different nodes), to hedge against the risk that any one copy of the process runs slowly or fails outright. According to my notes, speculative execution is a major part of NuoDB’ architecture as well.
I’ve rambled on for two long posts, which seems like plenty — but this survey is in no way complete. Other subjects I could have covered include but are hardly limited to:
- Occasionally-connected operation, which for example is a design point of CouchDB, SQL Anywhere (sort of), and most kinds of mobile business intelligence.
- Avoiding planned downtime — i.e., operating despite self-inflicted wounds.
- Data cleaning and master data management, both of which exist in large part to fix errors people have made in the past.
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:
- Database changes (inserts, deletes, updates) are applied expeditiously to disk. Furthermore …
- … a log is kept of the (instructions for) changes. If a change is detected as not going through, it can be reapplied.
- Data is often kept in multiple copies automagically, whether that is governed by the storage systems or the DBMS itself, with background resyncing if one copy is known to go bad.
- These ideas can be pushed further into:
- High availability — the database is mirrored onto storage that is controlled by a second server, which runs the same software as the primary.
- Disaster recovery — same idea as HA, but off-site to protect against site disasters. DR often cuts various corners, as compared to same-site HA, in the speed and/or quality of service with which the remote instance would take over from the primary site.
Valuable though they are, none of these techniques protects against data corruption that occurs via software errors or security breaches, because in those cases the system will likely do a great job of making incorrect changes consistently across all copies of the data. Hence there also needs to be a backup/restore capability, key aspects of which start:
- You periodically create and lock down snapshots of the database.
- In case of comprehensive failure, you load the most recent snapshot you trust, and roll forward by re-applying changes memorialized in the log. (Of course, you avoid re-applying spurious changes that should not have occurred.)
For many users, it is essential that backups be online/continuous, rather than requiring the database to periodically be taken down.
None of this is restricted to what in a relational database would be single-row or at least single-table changes. Transaction semantics cover the case that several changes must either all change or all fail together. Typically all the changes are made; the system observes that they’ve all been made; only then is a commit issued which makes the changes stick. One complexity of this approach is that you need a way to quickly undo changes that don’t get committed.
Locks, latches and timestamps
If a database has more than one user, two worrisome possibilities arise:
- Two conflicting changes might be attempted against the same data.
- One user might try to read data at the moment another user’s request was changing it.
So it is common to lock the portion of the database that is being changed. A major area of competition in the early 1990s — and a big contributor to Sybase’s decline — was the granularity of locks (finer-grained is better, with row-level locking being the relational DBMS gold standard).
In-memory locks are generally called latches; I don’t know what the other differences between locks and latches are.
Increasingly many DBMS are designed with an alternative approach, MVCC (Multi-Version Concurrency Control); indeed, I’m now startled when I encounter one that makes a different choice. The essence of MVCC is that, to each portion (e.g. row) of data, the system appends very granular metadata — a flag for validation/invalidation and/or a timestamp. Except for the flags, the data is never changed until a cleanup operation; rather, new data is written, with later timestamps and validating flags. The payoff is that, while it may still be necessary to lock data against the possibility of two essentially simultaneous writes, reads can proceed without locks, at the cost of a minuscule degree of freshness. For if an update is in progress that affects the data needed for a read, the read just targets versions of the data slightly earlier in time.
MVCC has various performance implications — writes can be faster because they are append-mainly, reads may be slower, and of course the cleanup is needed. These tradeoffs seem to work well even for the query-intensive workloads of analytic RDBMS, perhaps because MVCC fits well with their highly sequential approach to I/O.
A DBMS that truly ran only in RAM would lose data each time the power turned off. Hence a memory-centric DBMS will usually also have some strategy for persisting data.
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.
Personal note: When I interviewed for academic and think-tank jobs in 1981, my favorite interview speech was the one on utility regulation across different qualities-of-service in the face of uncertain supply and demand. I’m really going back to my roots here.
What I wrote in 2007 — and which garnered considerable discussion at the time — still applies:
Net neutrality is both necessary and workable for what I call Jeffersonet, which comprises the “classical”, bandwidth-light parts of the Internet. Thus, it includes e-mail, instant messaging, much e-commerce, and just about every website created in the first 13 or so years of the Web. Jeffersonet is the greatest tool in human history to communicate research, teaching, news, and political ideas, or to let tiny businesses compete worldwide. Any censorship of Jeffersonet – even if just of the self-interested large-enterprise commercial kind – would be a terrible loss. Net neutrality is workable for Jeffersonet because – well, because it’s already working just fine. Jeffersonet doesn’t need anything beyond current levels of bandwidth and reliability. So there’s no reason to mess with what’s working, other than simple profit-hungry greed.
Network neutrality opponents, however, point to evolving and future technologies, technically more demanding than what the current Internet can well support. Their uses are centered on what I call Edisonet – communication-rich applications such as entertainment, gaming, telephony, telemedicine, teleteaching, or telemeetings of all kinds. Reliable, tiered service is needed for these applications, and somebody has to pay for it. Even so – and this is a key point — the payment scheme should be as favorable to application-developer competition as possible.
So does what I wrote earlier this year:
I think the anti-discrimination argument for network neutrality has much merit. But I also think there are some kinds of payment structure that could leave the playing field fairly level. Imagine, if you will, that:
- Consumers are charged for data, speed of connection, reliability of delivery, or anything else, but …
- … internet companies have the ability to absorb those charges on consumers’ behalf, but can only do so …
- … one interaction at a time, with no volume discounts, via an automated system that is open to everybody.
Such a system is surely technologically feasible — indeed, it is at least as feasible as the online advertising networks that already exist. Further, it would be possible for the system to have nice features such as:
- Telcos could implement forms of peak load pricing, for those times when their network capacity actually is under stress.
- “Edge provider” internet companies could pay subsidies only on behalf of certain consumers, where those consumers are selected in all the complex ways that advertisements are currently targeted.
To see how this could look, let’s distinguish among some categories of market participant, and consider what kinds of business complexity they can reasonably be expected to endure.
- True consumers.Their situation probably should and will remain much as it is today:
- Simple choices of bundled connectivity-service plans.
- Consumption of particular sites and services on the basis of e-commerce, subscription or free (ad-supported or otherwise).
- Regulation needed to control — likely with partial success — the huge oligopolists who market, sell and supply the connectivity services.
- Other kinds of end customer(e.g. businesses acting as consumers).
- Most of their internet consumption is akin to true consumers’.
- In addition, they should be able to obtain business-critical services with SLA (Service Level Agreement) guarantees for reliability and speed.
- Connectivity providers. They can handle any kind of complexity their customers can, provided the equipment exists to deliver on the promises they make. This is true of customer-facing “last mile” and intermediate/wholesale telecommunication firms alike.
- Market-makers/middlemen for the new market(s) I’m suggesting be created. (Analogous to ad-tech real-time auctioneers/clearing houses.) It’s their job to handle the complexity everybody else needs.
- Publishers, e-commerce vendors and other internet-based enterprises – or, similarly, the internet parts of brick-and-mortar businesses.Here is where the analysis gets interesting.
- The reason we’re having this discussion is that certain large internet-based enterprises want the ability to buy SLA guarantees for their service delivery.
- If they’re able to do so, many of their competitors will suddenly develop that desire as well.
- Hence the business of SLA-guarantee purchasing needs to be organized not just for the benefit of a few large internet companies, but also to suit smaller companies who might wish to compete with them.
- What kind of SLA will a smaller/generic internet company want to be able to buy? In a nutshell, they’ll want to guarantee quality and reliability end-to-end on a customer-by-customer or interaction-by-interaction basis.
The distinction I’m drawing here is:
- Huge companies like Netflix or Google can buy their SLAs piecemeal — one deal to set up a private wholesale network, another deal to guarantee “last-mile” delivery, etc.
- But mom-and-pop web businesses don’t have that luxury. They need to buy delivery on an all-or-nothing basis.
Making the latter happen is, I maintain, just another job for the middlemen — provided those middlemen come into existence.
And so, net neutrality is easy to solve except for one chicken-egg problem:
- The right middlemen need to be created …
- … for a business that doesn’t yet exist …
- … and which indeed can’t exist unless they’re first created …
- … which nobody has a strong incentive to do unless the business is first shown to (be likely to) exist.
I hope somebody — perhaps an existing ad-tech company — gambles on setting up the needed clearinghouse, and gets richly rewarded for doing so.
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.
1. The recently-announced Cloudera/MongoDB relationship* is still at the Barney stage. That said, I’m optimistic that their stated intention to add substance to the relationship will eventually come to fruition. If nothing else, the two companies have high regard for each other, at least at the Mike Olson/Max Schireson level.
*That’s one of numerous deals with my fingerprints on it, but in this case only lightly. It was probably on track to happen even without my nudges.
2. Most of what I talked about when I visited MongoDB is confidential; the public stuff was mainly in my recent MongoDB technology post. But in one exception, I asked Max for an update as to MongoDB enterprise use cases. He reported a cluster in data combination, especially but not only in use cases which have both high-volume part and dynamic-schema aspects. Specific examples Max cited included:
- Tracking financial holdings from a variety of asset classes — especially if derivatives are involved, because they have a dynamic-schema aspect.
- Product catalogs, including for use on web sites.
- Customer information.
- Patient information.
3. I didn’t ask everybody I saw in California about business trends, and much of what we did discuss was confidential. That said:
- MapR was proud of its numbers.
- So was DataStax.
- ClearStory has a bunch of Very Big Enterprises as customers, mainly but not only in consumer sectors (e.g. retail, packaged goods).
4. Platfora is focusing a bit, starting with clickstream and security — i.e., event series stuff. And by the way, they report that the term “event series” is working well for them.
5. I gather from a variety of comments and conversations that Amazon Redshift has achieved considerable traction.
6. Something I can’t find evidence of having posted before: I think multiple businesses monitor online sales or similar business successes as a guide to network problems. eBay did this via a custom in-memory MOLAP (Multidimensional Online Analytic Process) system years ago. Best evidence that this is hardly restricted to eBay: all the “me-too” responses I get from telling that story.
7. Citus Data tells me that as of PostgreSQL 9.4, Postgres will be able to return just the part of a JSON column needed for a query. This is as opposed to storing the whole thing as text and only retrieving it in its entirety.
8. In the comments to my “Spark on fire” post, Patrick McFadin pointed out that Mahout is transitioning from MapReduce to Spark. (All new work will be on Spark, although old MapReduce-based routines will continue to be supported.) It turns out that Derrick Harris wrote about that over a month ago, and I just missed the news.
9. Also in predictive analytics — there are rumblings that R could eventually be supplanted by Julia, although R’s massive libraries of algorithms still give it the advantage now.
10. Multiple vendors, fed up with the intermittent slowdowns from garbage collection, are moving some processing off the Java heap. Unfortunately, I neglected to ask any of them what the remaining differences then were between Java and C++ programming.
11. And to finish on a light note: BDAS — the project of which Spark is only a part — is pronounced “bad-ass”, something I first heard from Dave Patterson.
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.
Notwithstanding what I said about “fat heads”, CitusDB does have a concept of Master nodes. These:
- Also use single-node copies of PostgreSQL.
- Are blessedly able to scale out, although their underlying databases are entirely replicated.
- Store no actual data, but do store metadata about each virtual node, including:
- Structural metadata.
- Min/max column values (for data skipping).
- But not (yet) stats to help with query optimization.
- Do some query planning and rewriting.
- Handle administration, some of which is nicely parallelized/centralized. (E.g., an index choice can be made once and automatically propagated across all the relevant virtual nodes.)
CitusDB is definitely in its early days. For example:
- If I understand correctly, the recent CitusDB 3.0 release is the first one on which data is redistributed among shards. Before that, you could only join tables that were either sharded on the same key, or else small enough to be broadcast-replicated across the whole cluster.
- SQL coverage isn’t great. (E.g., no Windowing.)
- Some hard-to-parallelize things aren’t implemented yet, e.g. exact median or generally-usable COUNT DISTINCT.
- ACID is still lacking. Writes are batch-only, micro-batch or otherwise as the case may be.
- CitusDB’s backup story is primitive, with the main options being:
- You can rely on having replicas on multiple nodes, even — if you like — in different data centers.
- You can backup each of the PostgreSQL nodes separately; CitusDB doesn’t yet offer automation for that.
- CitusDB’s query optimization sounds pretty primitive.
- I don’t recall Citus telling me of serious workload management.
- CitusDB compression is block-level only. (PostgreSQL’s version of Lempel-Ziv.)
Still, the Citus Data folks seem to have good ideas, including some — as yet undisclosed — plans going forward. So if it sounds as if CitusDB might fit your needs better than more established scale-out RDBMS do, I’d encourage you to take a look at what Citus offers.
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:
- Some MemSQL users are running 7- or 8-way joins and other long-ish SQL statements.
- But MemSQL doesn’t yet have fully peer-to-peer data redistribution.
- MemSQL “leaves” only talk to MemSQL “aggregator nodes,” not each other …
- … but note the plural on “aggregator nodes”, which should immunize MemSQL from the worst of “fat head” bottlenecks.
- Of course, you can sometimes get join locality by sharding multiple tables on the same key …
- … or by broadcast-replicating tables that are sufficiently small.
- Better SQL coverage — e.g. SQL Windowing — is coming soon.
- MemSQL believes it has an aggressive data skipping story.
- MemSQL doesn’t yet have a true workload management story; they’re still at the stage “Our queries run so fast not many of them have to be active at once, and if things nevertheless get too busy we have some throttling capabilities.” But MemSQL at least sounds aware of the difference between that and true workload management, which puts them ahead of some other vendors I talk with.
- MemSQL doesn’t have stored procedures. In particular, since MemSQL (the product) generates code on the fly, MemSQL (the company) doesn’t think the performance benefits of stored procedure pre-compilation are needed.
And finally, MemSQL’s column-store compression story — which I mangled in a previous post — goes like this:
- There are numerous compression algorithm choices, both columnar (e.g. dictionary/tokenization, run-length encoding) and block (Lempel-Ziv, I presume in multiple variations).
- Compression is block-by-block, something I hear more commonly these days than Vertica’s alternative of global compression choices.
- The choice of compression scheme is automagic for each block, unless you give explicit hints.
- Default block size for the columnar store is 10 million rows.
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.
Otherwise, I didn’t gain much new insight into actual flash uptake. Everybody thinks flash is or soon will be very important; but in many segments, folks are trading off disk vs. RAM without worrying much about the intermediate flash alternative.
I visited two Hadoop distribution vendors this trip, namely the ones who are my clients – Cloudera and MapR. I remembered to ask one of them, Cloudera, about typical Hadoop hardware, and got answers that sounded consistent with hardware trends Hortonworks told me about last August. The story is, more or less:
- The default assumption remains $20-30K/node, 2 sockets, 12 disks. (Edit: See lively price discussion in the comments below.)
- Most hardware vendors have standard/default Hadoop boxes by now, and in many cases customers just buy what’s on offer.
- The aforementioned disks sometimes get up to 4 terabytes now.
- 128GB is now the norm for RAM. 256GB is common. Higher amounts are seen, up to – in rare cases – 2-4 TB.
- Flash is of interest, but isn’t being demanded much yet. This could change when flash’s storage density matches disk’s.
- Flash interest is highest for Impala.
Cloudera suggested that the larger amounts of RAM tend to be used when customers frame the need as putting certain analytic datasets entirely in RAM. This rings true to me; there’s lots of evidence that users think that way, and not just in analytic cases. This is probably one of the reasons that they often jump straight from disk to RAM without fully exploring the opportunities of flash.
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,
- Cloudera will move rapidly to make its Hadoop distribution be a superset of Intel’s.
- When that is accomplished, Intel will get out of the Hadoop distribution business, and Cloudera will take over Intel’s Hadoop customers, which I still think are concentrated in Asia, especially China and India.
Notes on that include:
- All of the above is planned to take place by the end of 2014.
- The biggest case of Intel technology superseding/subsuming Cloudera’s will be security. (I.e., Cloudera Sentry will become a subset of Rhino.)
- Intel’s Hadoop engineers will remain Intel employees.
Medium-term Intel/Cloudera collaboration plans include looking for things to push down to the chip, such as:
I didn’t get into that in any detail with Cloudera, and am not sure I’m convinced yet much will come of it.
Finally, Cloudera and Intel will of course talk a lot about adapting to computer architecture opportunities and trends. I expect that part to go well, because Intel has strong relationships of that kind even with software companies it doesn’t own large percentages of.
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.
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.
In an unfortunate non-development, Tachyon is not (yet?) part of Spark, and so it is hard for a Spark job’s data to be shared with other jobs (Spark or otherwise) or processes. That said:
- The tight integration of data structures and processes gives similar performance benefits to those of in-process vs. out-of-process in-database analytic functions. (It also of course raises similar stability concerns, but those seem less important in the case of Spark than of a true DBMS.)
- From a Hadoop vendor’s standpoint, Tachyon’s benefit of not requiring HDFS (Hadoop Distributed File System) isn’t important, and Tachyon somewhat conflicts with a newish effort called HDFS Caching.
A couple of Spark machine learning stories are very cool, in that they involve intra-day retraining of models. The better-known one is Yahoo’s, which in a prototype built in 120 lines of code trains a new model for recommendation of each candidate top-page news story. When I challenged that anecdote, Ion told me about his own former company Conviva, which retrains models every minute to decide which particular source of streaming video each client system will be connected to.
I am generally skeptical of immature SQL efforts, and SparkSQL is no exception. That said, it seems to be going in sensible directions, which should be welcome to those folks who used or were planning to use Shark anyway.
- SparkSQL actually has its own optimizer, rather than using the inappropriate Hive one. As with many new optimizers, it’s starting out rule-based, but is planned to become cost-based down the road.
- SparkSQL can run queries against data that’s either inside Spark or outside-but-accessible.
- SparkSQL can be accessed via Python and other APIs.
- Spark works with the Hive metastore, nee’ HCatalog.
And finally, there’s no public news as to what Databricks’ own business is. I think that’s a bit silly, but in fairness:
- The Spark 1.0 launch will consume every bit of marketing bandwidth they have.
- They don’t yet want to commit to a delivery date of their first offering.
When I’m asked to talk to academics, the requested subject is usually a version of “What should we know about what’s happening in the actual market/real world?” I then try to figure out what the scholars could stand to hear that they perhaps don’t already know.
In the current case (Berkeley next Tuesday), I’m using the title “Necessary complexity”. I actually mean three different but related things by that, namely:
- No matter how cool an improvement you have in some particular area of technology, it’s not very useful until you add a whole bunch of me-too features and capabilities as well.
- Even beyond that, however, the simple(r) stuff has already been built. Most new opportunities are in the creation of complex integrated stacks, in part because …
- … users are doing ever more complex things.
While everybody on some level already knows all this, I think it bears calling out even so.
I previously encapsulated the first point in the cardinal rules of DBMS development:
Rule 1: Developing a good DBMS requires 5-7 years and tens of millions of dollars.
That’s if things go extremely well.
Rule 2: You aren’t an exception to Rule 1.
- Concurrent workloads benchmarked in the lab are poor predictors of concurrent performance in real life.
- Mixed workload management is harder than you’re assuming it is.
- Those minor edge cases in which your Version 1 product works poorly aren’t minor after all.
My recent post about MongoDB is just one example of same.
Examples of the second point include but are hardly limited to:
- Hadoop and its ecosystem.
- The general trend of supporting multiple data paradigms in one system …
- … sometimes via schema-on-need.
- DBMS vendors’ work to exploit multiple kinds of storage in one system, from Microsoft to MemSQL.
- WibiData and Kiji.
BDAS and Spark make a splendid example as well.
As to the third point:
- In ever more use cases, the essential simplicity of the relational data model is fundamentally obsolete.
- It’s been generally accepted for a couple of years that analytic data topologies often need to be complex.
- Predictive models are getting markedly more complex as well.
Bottom line: Serious software has been built for over 50 years. Very little of it is simple any more.
I caught up with my clients at MongoDB to discuss the recent MongoDB 2.6, along with some new statements of direction. The biggest takeaway is that the MongoDB product, along with the associated MMS (MongoDB Management Service), is growing up. Aspects include:
- An actual automation and management user interface, as opposed to the current management style, which is almost entirely via scripts (except for the monitoring UI).
- That’s scheduled for public beta in May, and general availability later this year.
- It will include some kind of integrated provisioning with VMware, OpenStack, et al.
- One goal is to let you apply database changes, software upgrades, etc. without taking the cluster down.
- A reasonable backup strategy.
- A snapshot copy is made of the database.
- A copy of the log is streamed somewhere.
- Periodically — the default seems to be 6 hours — the log is applied to create a new current snapshot.
- For point-in-time recovery, you take the last snapshot prior to the point, and roll forward to the desired point.
- A reasonable locking strategy!
- Document-level locking is all-but-promised for MongoDB 2.8.
- That means what it sounds like. (I mention this because sometimes an XML database winds up being one big document, which leads to confusing conversations about what’s going on.)
- Security. My eyes glaze over at the details, but several major buzzwords have been checked off.
- A general code rewrite to allow for (more) rapid addition of future features.
Of course, when a DBMS vendor rewrites its code, that’s a multi-year process. (I think of it at Oracle as spanning 6 years and 2 main-number releases.) With that caveat, the MongoDB rewrite story is something like:
- Updating has been reworked. Most of the benefits are coming later.
- Query optimization and execution have been reworked. Most of the benefits are coming later, except that …
- … you can now directly filter on multiple indexes in one query; previously you could only simulate doing that by pre-building a compound index.
- One of those future benefits is more index types, for example R-trees or inverted lists.
- Concurrency improvements are down the road.
- So are rewrites of the storage layer, including the introduction of compression.
Also, you can now straightforwardly transform data in a MongoDB database and write it into new datasets, something that evidently wasn’t easy to do before.
One thing that MongoDB is not doing is offer any ODBC/JDBC or other SQL interfaces. Rather, there’s some other API — I don’t know the details — whereby business intelligence tools or other systems can extract views, and a few BI vendors evidently are doing just that. In particular, MicroStrategy and QlikView were named, as well as a couple of open source usual-suspects.
As of 2.6, MongoDB seems to have a basic integrated text search capability — which however does not rise to the search functionality level that was in Oracle 7.3.2. In particular:
- 15 Western languages are supported with stopwords, tokenization, etc.
- Search predicates can be mixed into MongoDB queries.
- The search language isn’t very rich; for example, it lacks WHERE NEAR semantics.
- You can’t tweak the lexicon yourself.
And finally, some business and pricing notes:
- Two big aspects of the paid-versus-free version of MongoDB (the product line) are:
- Management tools.
- Well, actually, you can get the management tools for free, but only on a SaaS basis from MongoDB (the company).
- If you want them on premises or in your part of the cloud, you need to pay.
- If you want MongoDB (the company) to maintain your backups for you, you need to pay.
- Customer counts include:
- At least 1000 or so subscribers (counting by organization).
- Over 500 (additional?) customers for remote backup.
- 30 of the Fortune 100.
And finally, MongoDB did something many companies should, which is aggregate user success stories for which they may not be allowed to publish full details. Tidbits include:
- Over 100 organizations run clusters with more than 100 nodes. Some clusters exceed 1,000 nodes.
- Many clusters deliver hundreds of thousands of operations per second (combined read and write).
- MongoDB clusters routinely store hundreds of terabytes, and some store multiple petabytes of data. Over 150 clusters exceed 1 billion documents in size. Many manage more than 100 billion documents.
If the makers of MMO RPGs (Massive Multi-Player Online Role-Playing Games) aren’t quite the worst database application developers in the world, they’re at least on the short list for consideration. The makers of Guild Wars didn’t even try to have decent database functionality. A decade later, when they introduced Guild Wars 2, the database-oriented functionality (auction house, real-money store, etc.) would crash for days at a time. Lord of the Rings Online evidently had multiple issues with database functionality. Now I’m playing Elder Scrolls Online, which on the whole is a great game, but which may have the most database screw-ups of all.
ESO has been live for less than 3 weeks, and in that time:
2. Guild functionality has at times been taken down while the rest of the game functioned.
3. Those problems aside, bank and guild bank functionality are broken, via what might be considered performance bugs. Problems I repeatedly encounter include:
- If you deposit a few items, the bank soon goes into a wait state where you can’t use it for a minute or more.
- Similarly, when you try to access a guild — i.e. group — bank, you often find it in an unresponsive state.
- If you make a series of updates a second apart, the game tells you you’re doing things too quickly, and insists that you slow down a lot.
- Items that are supposed to “stack” appear in 2 or more stacks; i.e., a very simple kind of aggregation is failing. There are also several other related recurring errors, which I conjecture have the same underlying cause.
In general, it seems like that what should be a collection of database records is really just a list, parsed each time an update occurs, periodically flushed in its entirety to disk, with all the performance problems you’d expect from that kind of choice.
4. Even stupider are the in-game stores, where fictional items are sold for fictional money. They have an e-commerce interface that is literally 15+ years out of date — items are listed with VERY few filtering options, and there is no way to change the sort. But even that super-primitive interface doesn’t work; in particular, filter queries frequently return incorrect empty-set responses.
5. Much as in other games, over 10 minutes of state changes can be lost.
Except perhaps for #5, these are all functions that surely are only loosely coupled to the rest of the game. Hence the other difficulties of game scaling and performance should have no bearing on them. Hence there’s no excuse for doing such a terrible job of development on large portions of gameplay functionality.
Based on job listings, ESO developer Zenimax doesn’t see database functionality as a major area to fix. This makes me sad.