Whenever somebody asks for my help on application technology strategy, I start by trying to ascertain three things. The absolute first is actually a prerequisite to almost any kind of useful conversation, which is to ascertain in general terms what the hell it is that we are talking about.
My second goal is to ascertain technology constraints. Three common types are:
- Compatible with legacy systems and/or enterprise standards.
- Cheap, free and/or open source.
- Proven, vetted by sufficiently many references, and/or generally having an “enterprise-y” reputation.
That’s often a short and straightforward discussion, except in those awkward situations when all three of my bullet points above are applicable at once.
The third item is usually more interesting. I try to figure out what is to be accomplished. That’s usually not a simple matter, because the initial list of goals and requirements is almost never accurate. It’s actually more common that I have to tell somebody to be more ambitious than that I need to rein them in.
Commonly overlooked needs include:
- If you want to sell something and have happy users, you need a good UI.
- You will also soon need tools and a UI for administration.
- Customers demand low-latency/fresh data. Your explanation of why they don’t really need it doesn’t contradict the fact that they want it.
- Providing data access and saying “You can hook up any BI tool you want and build charts” is not generally regarded as offering a good UI.
- When “adding analytics” to something previously focused on short-request processing, it is common to underestimate the variety of things users will soon want to do. (One common reason for this under-estimate is that after years of being told it can’t be done, they’ve learned not to ask.)
And if you take one thing away from this post, then take this:
- If you “know” exactly which features are or aren’t helpful to users, …
- .. and if you supply only what you “know” they should use, …
- … then you will discover that what you “knew” wasn’t really accurate.
I guarantee it.
So far what I’ve said can be summarized as “Figure out what you’re trying to do, and what constraints there are on your choices for doing it.” The natural next step is to list the better-thought-of choices that meet your constraints, and — voila! — you have a short list. That’s basically correct, but there’s one significant complication.
Speaking of complications, what I’m portraying as a kind of linear/waterfall decision process of course usually involves lots of iteration, meandering around and general wheel-spinning. Real life is messy.
Simply put, there are many different kinds of application project. Other folks’ experience may not be as applicable to your case as you hope, because your case is different. So the rest of this post contains a checklist of distinctions among various different kinds of application project.
For starters, there are at least two major kind(s) of software development.
- Many projects fit the traditional development model, elements of which are:
- You — and this is very much a plural “you” — code something up more or less from scratch, using whatever language(s) and/or framework(s) you think make sense.
- You break the main project into pieces in obvious ways (e.g. server back end vs. mobile front), and then into further pieces for manageability.
- There may also be database designs, test harnesses, connectors to other apps and so on.
- But there are many other projects in which smaller bits of configuration and/or scripting are the essence of what you do.
- This is particularly common in analytics, where there might be business intelligence tools, ETL tools, scripts running against Hadoop and so on. The original building of a data warehouse/hub/lake/reservoir may also fit this model.
- It’s also what you do to get a major purchased packaged application into actual production.
- It also is often what happens for websites that serve “content”.
Other significant distinctions include:
- In-house vs. software-for-resale. If the developing organization is handing code to somebody else, then we’re probably talking about a more traditional kind of project. But if the whole thing is growing organically in-house, the script-spaghetti alternative may well be viable (in those projects for which it seems appropriate). Important subsidiary distinctions start with:
- (If in-house) Truly in-house vs. out-sourced.
- (If for resale) On-premises vs. SaaS. Or maybe not.
- Kind(s) of analytics, if any. Technologies and development processes used can be very different depending upon whether the application features:
- Business intelligence (not particularly real-time) as its essence.
- Reporting or other BI as added functionality to an essentially operational app.
- Low-latency BI, perhaps supported by (other) short-request processing.
- Predictive model scoring.
- The role(s) of the user(s). This influences how appealing and easy the UI needs to be.* Requirements are very different, for example, among:
- Classic consumer-facing websites, with recommenders and so on.
- Marketing websites targeted at a small group of business-to-business customers.
- Data-sharing websites for existing consumer stakeholders.
- Cheery benefits-information websites that the HR department wants employees to look at.
- Purely internal apps meant to be used by (self-)important executives.
- Internal apps meant to be used by line workers who will be given substantial training on them.
- Certain kinds of application project stand almost separately from the rest of these considerations, because their starting point is legacy apps. Examples may be found among:
- Migration/consolidation projects.
- Refactoring projects.
- Addition of incremental functionality.
*It also influences security, all good practices for securing internal apps notwithstanding.
Much also depends on the size and sophistication of the organization. What the “organization” is depends a bit on context:
- In the case of software products, SaaS (Software as a Service) or other internet services, it is primarily the vendor. However …
- … in B2B cases the sophistication of the customer organizations can also matter.
- In the case of in-house enterprise development, there’s only one enterprise involved (duh). However …
- … the “department” vs. “IT” distinction may be very important.
Specific considerations of this kind start:
- Is me-too functionality enough, or does the enterprise seek competitive advantage through technology?
- What kinds of technical risk does it seem prudent and desirable to take?
And that, in a nutshell, is why strategizing about application technology is often more complicated than it first appears.
- My November, 2015 post on issues in enterprise application software links to a number of other relevant posts.
- One of those (the same month) briefly surveyed actual choices in technology support for enterprise apps.
- A number of my posts draw distinction among different analytic use cases. An April, 2015 example points to some of the earlier ones.
- My July, 2012 categorization of kinds of BI is particularly relevant.
- A November, 2012 post focused on assessing the supposed need for speed.
- My September, 2011 strategic worksheet is evergreen.
In a companion introduction to Kafka post, I observed that Kafka at its core is remarkably simple. Confluent offers a marchitecture diagram that illustrates what else is on offer, about which I’ll note:
- The red boxes — “Ops Dashboard” and “Data Flow Audit” — are the initial closed-source part. No surprise that they sound like management tools; that’s the traditional place for closed source add-ons to start.
- “Schema Management”
- Is used to define fields and so on.
- Is not equivalent to what is ordinarily meant by schema validation, in that …
- … it allows schemas to change, but puts constraints on which changes are allowed.
- Is done in plug-ins that live with the producer or consumer of data.
- Is based on the Hadoop-oriented file format Avro.
Kafka offers little in the way of analytic data transformation and the like. Hence, it’s commonly used with companion products.
- Per Confluent/Kafka honcho Jay Kreps, the companion is generally Spark Streaming, Storm or Samza, in declining order of popularity, with Samza running a distant third.
- Jay estimates that there’s such a companion product at around 50% of Kafka installations.
- Conversely, Jay estimates that around 80% of Spark Streaming, Storm or Samza users also use Kafka. On the one hand, that sounds high to me; on the other, I can’t quickly name a counterexample, unless Storm originator Twitter is one such.
- Jay’s views on the Storm/Spark comparison include:
- Storm is more mature than Spark Streaming, which makes sense given their histories.
- Storm’s distributed processing capabilities are more questionable than Spark Streaming’s.
- Spark Streaming is generally used by folks in the heavily overlapping categories of:
- Spark users.
- Analytics types.
- People who need to share stuff between the batch and stream processing worlds.
- Storm is generally used by people coding up more operational apps.
If we recognize that Jay’s interests are obviously streaming-centric, this distinction maps pretty well to the three use cases Cloudera recently called out.
Complicating this discussion further is Confluent 2.1, which is expected late this quarter. Confluent 2.1 will include, among other things, a stream processing layer that works differently from any of the alternatives I cited, in that:
- It’s a library running in client applications that can interrogate the core Kafka server, rather than …
- … a separate thing running on a separate cluster.
The library will do joins, aggregations and so on, and while relying on core Kafka for information about process health and the like. Jay sees this as more of a competitor to Storm in operational use cases than to Spark Streaming in analytic ones.
We didn’t discuss other Confluent 2.1 features much, and frankly they all sounded to me like items from the “You mean you didn’t have that already??” list any young product has.
- My October, 2014 post on Streaming for Hadoop is a sort of predecessor to this two-post series.
- Kafka has gotten considerable attention and adoption in streaming.
- Kafka is open source, out of LinkedIn.
- Folks who built it there, led by Jay Kreps, now have a company called Confluent.
- Confluent seems to be pursuing a fairly standard open source business model around Kafka.
- Confluent seems to be in the low to mid teens in paying customers.
- Confluent believes 1000s of Kafka clusters are in production.
- Confluent reports 40 employees and $31 million raised.
At its core Kafka is very simple:
- Kafka accepts streams of data in substantially any format, and then streams the data back out, potentially in a highly parallel way.
- Any producer or consumer of data can connect to Kafka, via what can reasonably be called a publish/subscribe model.
- Kafka handles various issues of scaling, load balancing, fault tolerance and so on.
So it seems fair to say:
- Kafka offers the benefits of hub vs. point-to-point connectivity.
- Kafka acts like a kind of switch, in the telecom sense. (However, this is probably not a very useful metaphor in practice.)
Jay also views Kafka as something like a file system. Kafka doesn’t actually have a file-system-like interface for managing streams, but he acknowledges that as a need and presumably a roadmap item.
The most noteworthy technical point for me was that Kafka persists data, for reasons of buffering, fault-tolerance and the like. The duration of the persistence is configurable, and can be different for different feeds, with two main options:
- Guaranteed to have the last update of anything.
- Complete for the past N days.
Jay thinks this is a major difference vs. messaging systems that have come before. As you might expect, given that data arrives in timestamp order and then hangs around for a while:
- Kafka can offer strong guarantees of delivering data in the correct order.
- Persisted data is automagically broken up into partitions.
Technical tidbits include:
- Data is generally fresh to within 1.5 milliseconds.
- 100s of MB/sec/server is claimed. I didn’t ask how big a server was.
- LinkedIn runs >1 trillion messages/day through Kafka.
- Others in that throughput range include but are not limited to Microsoft and Netflix.
- A message is commonly 1 KB or less.
- At a guesstimate, 50%ish of messages are in Avro. JSON is another frequent format.
Jay’s answer to any concern about performance overhead for current or future features is usually to point out that anything other than the most basic functionality:
- Runs in different processes from core Kafka …
- … if it doesn’t actually run on a different cluster.
For example, connectors have their own pools of processes.
I asked the natural open source question about who contributes what to the Apache Kafka project. Jay’s quick answers were:
- Perhaps 80% of Kafka code comes from Confluent.
- LinkedIn has contributed most of the rest.
- However, as is typical in open source, the general community has contributed some connectors.
- The general community also contributes “esoteric” bug fixes, which Jay regards as evidence that Kafka is in demanding production use.
Jay has a rather erudite and wry approach to naming and so on.
- Kafka got its name because it was replacing something he regarded as Kafkaesque. OK.
- Samza is an associated project that has something to do with transformations. Good name. (The central character of The Metamorphosis was Gregor Samsa, and the opening sentence of the story mentions a transformation.)
- In his short book about logs, Jay has a picture caption “ETL in Ancient Greece. Not much has changed.” The picture appears to be of Sisyphus. I love it.
- I still don’t know why he named a key-value store Voldemort. Perhaps it was something not to be spoken of.
What he and his team do not yet have is a clear name for their product category. Difficulties in naming include:
- Kafka is limited and simple. But of course Confluent has plans to broaden its capabilities.
- It’s long been hard to decide whether to talk about “events”, “streams” or both.
- “Streaming” has another tech meaning, in the context of video, songs, etc.
- One candidate name, “event hub”, has already been grabbed by IBM and Microsoft for their specific offerings.
- Naming is always hard in general.
Confluent seems to be using “stream data platform” as a placeholder. As per the link above, I once suggested Data Stream Management System, or more concisely Datastream Manager. “Event”, “event stream” or “event series” could perhaps be mixed in as well. I don’t really have an opinion yet, and probably won’t until I’ve studied the space in a little more detail.
And on that note, I’ll end this post for reasons of length, and discuss Kafka-related technology separately.
- My October, 2014 post on Streaming for Hadoop is a sort of predecessor to this two-post series.
Cloudera released Version 2 of Cloudera Director, which is a companion product to Cloudera Manager focused specifically on the cloud. This led to a discussion about — you guessed it! — Cloudera and the cloud.
Making Cloudera run in the cloud has three major aspects:
- Cloudera’s usual software, ported to run on the cloud platform(s).
- Cloudera Director, which for example launches cloud instances.
- Points of integration, e.g. taking information about security-oriented roles from the platform and feeding then to the role-based security that is specific to Cloudera Enterprise.
Features new in this week’s release of Cloudera Director include:
- An API for job submission.
- Support for spot and preemptable instances.
- High availability.
- Some cluster repair.
- Some cluster cloning.
I.e., we’re talking about some pretty basic/checklist kinds of things. Cloudera Director is evidently working for Amazon AWS and Google GCP, and planned for Windows Azure, VMware and OpenStack.
As for porting, let me start by noting:
- Shared-nothing analytic systems, RDBMS and Hadoop alike, run much better in the cloud than they used to.
- Even so, it seems that the future of Hadoop in the cloud is to rely on object storage, such as Amazon S3.
That makes sense in part because:
- The applications where shared nothing most drastically outshines object storage are probably the ones in which data can just be filtered from disk — spinning-rust or solid-state as the case may be — and processed in place.
- By way of contrast, if data is being redistributed a lot then the shared nothing benefit applies to a much smaller fraction of the overall workload.
- The latter group of apps are probably the harder ones to optimize for.
But while it makes sense, much of what’s hardest about the ports involves the move to object storage. The status of that is roughly:
- Cloudera already has a lot of its software running on Amazon S3, with Impala/Parquet in beta.
- Object storage integration for Windows Azure is “in progress”.
- Object storage integration for Google GCP it is “to be determined”.
- Security for object storage — e.g. encryption — is a work in progress.
- Cloudera Navigator for object storage is a roadmap item.
When I asked about particularly hard parts of porting to object storage, I got three specifics. Two of them sounded like challenges around having less detailed control, specifically in the area of consistency model and capacity planning. The third I frankly didn’t understand,* which was the semantics of move operations, relating to the fact that they were constant time in HDFS, but linear in size on object stores.
*It’s rarely obvious to me why something is o(1) until it is explained to me.
Naturally, we talked about competition, differentiation, adoption and all that stuff. Highlights included:
- In general, Cloudera’s three big marketing messages these days can be summarized as “Fast”, “Easy”, and “Secure”.
- Notwithstanding the differences as to which parts of the Cloudera stack run on premises, on Amazon AWS, on Microsoft Azure or on Google GCP, Cloudera thinks it’s important that its offering is the “same” on all platforms, which allows “hybrid” deployment.
- In general, Cloudera still sees Hortonworks as a much bigger competitor than MapR or IBM.
- Cloudera fondly believes that Cloudera Manager is a significant competitive advantage vs. Ambari. (This would presumably be part of the “Easy” claim.)
- In particular, Cloudera asserts it has better troubleshooting/monitoring than the cloud alternatives do, because of superior drilldown into details.
- Cloudera’s big competitor on the Amazon platform is Elastic MapReduce (EMR). Cloudera points out that EMR lacks various capabilities that are in the Cloudera stack. Of course, versions of these capabilities are sometimes found in other Amazon offerings, such as Redshift.
- Cloudera’s big competitor on Azure is HDInsight. Cloudera sells against that via:
- General Cloudera vs. Hortonworks distinctions.
Cloudera also offered a distinction among three types of workload:
- ETL (Extract/Transform/Load) and “modeling” (by which Cloudera seems to mean predictive modeling).
- Cloudera pitches this as batch work.
- Cloudera tries to deposition competitors as being good mainly at these kinds of jobs.
- This can be reasonably said to be the original sweet spot of Hadoop and MapReduce — which fits with Cloudera’s attempt to portray competitors as technical laggards.
- Cloudera observes that these workloads tend to call for “transient” jobs. Lazier marketers might trot out the word “elasticity”.
- BI (Business Intelligence) and “analytics”, by which Cloudera seems to mainly mean Impala and Spark.
- “Application delivery”, by which Cloudera means operational stuff that can’t be allowed to go down. Presumably, this is a rough match to what I — and by now a lot of other folks as well — call short-request processing.
While I don’t agree with terminology that says modeling is not analytics, the basic distinction being drawn here make considerable sense.
I’m on two overlapping posting kicks, namely “lessons from the past” and “stuff I keep saying so might as well also write down”. My recent piece on Oracle as the new IBM is an example of both themes. In this post, another example, I’d like to memorialize some points I keep making about business intelligence and other analytics. In particular:
- BI relies on strong data access capabilities. This is always true. Duh.
- Therefore, BI and other analytics vendors commonly reinvent the data management wheel. This trend ebbs and flows with technology cycles.
Similarly, BI has often been tied to data integration/ETL (Extract/Transform/Load) functionality.* But I won’t address that subject further at this time.
*In the Hadoop/Spark era, that’s even truer of other analytics than it is of BI.
My top historical examples include:
- The 1970s analytic fourth-generation languages (RAMIS, NOMAD, FOCUS, et al.) commonly combined reporting and data management.
- The best BI visualization technology of the 1980s, Executive Information Systems (EIS), was generally unsuccessful. The core reason was a lack of what we’d now call drilldown. Not coincidentally, EIS vendors — notably leader Comshare — didn’t do well at DBMS-like technology.
- Business Objects, one of the pioneers of the modern BI product category, rose in large part on the strength of its “semantic layer” technology. (If you don’t know what that is, you can imagine it as a kind of virtual data warehouse modest enough in its ambitions to actually be workable.)
- Cognos, the other pioneer of modern BI, depending on capabilities for which it needed a bundled MOLAP (Multidimensional OnLine Analytic Processing) engine.
- But Cognos later stopped needing that engine, which underscores my point about technology ebbing and flowing.
I’m not as familiar with the details for MicroStrategy, but I do know that it generates famously complex SQL so as to compensate for the inadequacies of some DBMS, which had the paradoxical effect of creating performance challenges for MicroStrategy used over more capable analytic DBMS, which in turn led at least Teradata to do special work to optimize MicroStrategy processing. Again, ebbs and flows.
More recent examples of serious DBMS-like processing in BI offerings may be found in QlikView, Zoomdata, Platfora, ClearStory, Metamarkets and others. That some of those are SaaS (Software as a Service) doesn’t undermine the general point, because in each case they have significant data processing technology that lies strictly between the visualization and data store layers.
- Context for this post may be found in my piece on The two sides of BI. (August, 2013)