Cubegeek

Off the Cuff on Gartner's 2014 BI Quadrant

Here at Cubegeek as well as at my other blogs I try to be as informal and devoid of bullshit as possible, and of course that includes the fact that I acknowledge the existence of bullshit in the necessity to say such an unpleasant word. But it's also possible that I, through the limited experiences of a generation or two in the software business do my share of promulgating opinions that are full of it. So be warned, here are some off the cuff remarks about the latest Gartner BI Quadrant informed by my couple dozen years in the business.

First off on Gartner itself. These guys are reliable and as long as I've been around, this magic quadrant assessment has been accurate and useful. They are marketing folks that understand what corporate buyers are looking for and what they are actually buying. They can't tell you very well how well the products work when it comes to implementation, or eventual TCO from a developer's point of view, but they are drivers of the checkboxes that matter in enterprise sales. In terms of understanding the market, you can't go too wrong reading Gartner. That said, only a longitudinal look at Gartner's stuff will give you an industry wide emphasis on what matters in the long term. As I read this (and I haven't for a while) I see how accurate it is to notice that Scorecards have dropped off the map. It was always a huge deal that brought tons of expertise, lots of controversy and little to no revenue. 

Anyway. Having been in post-sales as well as pre-sales consulting I get to understand marketing and development as well has how different organizations approach and deal with BI tech & products. Lucky me. 

Let's start at the low end. I don't know a lot of these vendors but I'll try to be blunt.

Infor used to have great marketing and a captive audience. They were very sophisticated about a few things and you could never separate their customers from them. That is a kind of death spiral. I'm not surprised. I think they were very closed.

Arcplan was a brilliant product 20 years ago. They innovated in visual programming back in the days of Powerbuilder. Their product was large, sophisticated, comprehensive and German - like a plane with a million switches in the cockpit, you could do anything. However most of the time you just wanted to do one or two things and it was very clumsy. Big learning curve, small payoff.

Actuate was a very nice small limited capability product run buy one or two guys, if I remember correctly. It was a low price alternative doing 80/20 on 70% of what big clients needed. I don't think they have enough revenue to extend the product much. 

Pentaho and Jaspersoft I'm going to put in the same bucket. These are low / no-cost open source community products with pay enterprise editions. I like the technical vision of the guys at Pentaho and they are very well marketed. Prior to coming out with that set of tools, the developers there made a standout product which was my favorite in the Hyperion universe. Back in the day when fat clients ruled, they were absolute masters. I haven't seen their latest stuff, and so I'm a little surprised to see them ranked equally with Jaspersoft. Pentaho has done a very admirable job understanding the dynamics of DW building and because of that their ETL vision is pretty fine. I trust them to get data from end to end elegantly.

Jaspersoft has made one of the nicest plug and play adhoc reporting front ends in the business with a zillion capabilities in the back end. They have some old heads in the BI field working with geeky java devs who, like a swarm of ants, have put everything into an Eclipse metaphor, and slowly added every little function possible with every open source widget. There's an upside and a downside to this. The internals of Jasper are unapproachable to any but java geeks and consequently the higher level UI design for application builders is unnecessarily complex. On the other hand, their approach to data in the domain concept, is very nice.

I haven't seen anything new from Panorama at all, so I hesitate to comment. They are deeply smart on database technology having made SQL Server into that thing that is MSAS. I would seriously consider whatever they do. I'm tempted to go to their website and eyeball their product offerings, but I would rather remain off the cuff.

Good Data should have been higher if their products were as good as their marketing. I think it's rather telling that they are not given the size of their company and market. So I suspect the product line is getting large and confusing without techical vision. Just a suspicion. 

Logi Analytics. Don't know them. Would check them out. Birst, same thing. IBI don't trust them. Never liked their technology. Same with Microstrategy. They built a product that was ridiculously complex and made their 20 somethings work 60 hour weeks to make it work - then depended on lock in and Moore's Law to save their bacon. They have too much money to be stupid, but elegance is something you can't buy and Microstrategy doesn't have it. Again. See how tightly this group is situated. There are no breakout products here. The companies on the large side cannot innovate much for fear of losing embedded customers - the companies on the small side cannot innovate much because that costs money.

Oracle does one thing better than anybody and that is provide every good support for a large range of very capable and diverse products. And whatever they can't build, they buy. However, survival of the fittest means no evolutionary leaps. They buy those competing products and let them sink or swim in an ever-expanding product line. Machiavellian, sure. It does work. 

Tibco. I don't understand them at all. Spotfire was supposed to be awesome, a simple capable product with a very sweet look and brain dead simple implementation. Methinks that mantle has been taken over by Qlikview, which makes it a market leader because they do 80/20 best. In fact, the only problem Qlikview has is that it operates on the same premise as Crystal Reports. It makes so many tasks so easy that it frustrates database experts by putting lots of development in the hands of non-developers. Three years ago they went big to a lot of people saying 'Its about time'. I suspect Qlik will be eating Tibco's lunch as time goes on, and those simpler products go the way of Actuate.

IBM, SAS, SAP
These are you sure bets. When it comes to BI, they tend to be a bit more focused than Oracle. That's my perception. SAP owns what used to be Business Objects and Sybase. Great combination, good integration, excellent rebranding and as always, steamroller implementation. SAP owns its customers like nobody else. SAS on the other hand loves its customers like nobody else and wraps them in a blanket of competence that goes across every platform. It's a good safe place to be. IBM has acquired what used to be Cognos, and is broadminded and competent enough on so very many levels as to not be threatened by anything or anybody. What's new about this is that I have sensed that what they are now talking about with data governance and smarter planet is going to pay off big time in the vision department. Just don't let them do the implementation. 

OK Microsoft. I honestly don't know what MSFT is doing these days that puts them so high up there. Obviously they have the resources to deliver. Obviously SQL Server is a workhorse that handles BI sized volumes of data capably. But it's just as obvious to me that as soon as they say something like Office integration or Sharepoint integration, I feel a hairball coming up. So I cheated and looked at PowerBI, and it looks like Windows 8 Sharepoint integration. I only peeked for 30 seconds mind you, and that's my sense of it. I have a strong bias against the Microsoft development environment that comes from years of frustration. There's no denying two things about Microsoft. One they put hundreds of millions of people into desktop computing. Two, they are not innovators when it comes to the web, in fact they suck pretty bad. So I have a fundamental difficulty with their core technologies and product philosophies, IE lack of Linux. So just really.. their interoperability path is frightening, and I really just can't get over that. All that said, somebody at Microsoft has been very busy.

Which brings us to the undisputed champions, Tableau. I saw this one coming, which is probably a serious reason why I decided to write this. Tableau as a company is full of shiny happy people and they have every reason to be so. They basically invented a visual language, and as an old Xerox head I like people who take visualization seriously. There are not a bunch of widgets slammed together using libraries anybody can find on the web, it's original stuff. Tableau has been marketed well, and they dominate the space occupied by marketing people - they've opened up doors for the kinds of databases I want to work with (columnar stores) and they're cloud friendly. But the bottom line is the capability of the tool, which is extraordinary and comprehensive. Everything about Tableau just screams 'good design'. 

Yes, I am about to get deep into Tableau this year, which is something I've kind of been saying more or less for several years. I don't see any reason why, as a BI front-end, Tableau can't rule for a long time, and this week I'll be setting up my own lab in AWS to start seeing what I can make that puppy do with Redshift. I suspect the answer will be a hell of a lot, depending on its cacheing strategy.  So I'll be debunking my own hype as time goes by, and I'll probably identify its fundamental flaw rather quickly - which will be that it runs only on Windows. But in the 64 bit world, Windows can be quite capable as an application server, so long as the stack is not extraordinarily thick. So yeah I am admitting to a certain degree that horizontal scaling of Tableau data servers may not be troublesome even though it's Windows based. We'll see. 

So there are my grains of salt to be reconciled with what Gartner actually says. There may be a lot of activity in the BI business world, but mastery of several certain principles is all it really takes. In the end, we are limited by human cognition and data quality, so these tools can only be so good. I put the premium on elegance of design, deployability and scalability. There's a lot in elegance of design that's very difficult to do, which is why certain new products can dominate this area quickly. So there's my gut, off the cuff. Best of luck with your decisions.

Errorfication

One of the great things about columnar databases is that you really don't have to worry about the cost of doing incremental updates to columns. In this post I'm going to talk about one of the ways we do that, and specifically about what I call 'errorfication'. 

Errorfication is a way of thinking about the gaps in your analytic data. Imagine that you have - oh lets's say six or seven fact and dimension tables that you are combining in new ways that have never been thought of before. It may sound surprising, but one day you should go and look at a big Teradata installation. I have. Chances are, Teradata being what it is, that there will be hundreds of tables that only exist there and nobody has really had time to experiment with what various kinds of multidimensional analysis can be done. Teradata developers and hardware and all of that is rather expensive, so you keep everything (one you've paid for a node, why not?) and you build what you must. Well we came into a situation like that and pulled off all the history for a half dozen tables or so and dropped it into S3 with a Producer. In our case, we chunked the data into one month segments which gave us all we needed. Some of the tables gave us good sized chunks, maybe a half a million records a month, others just a few thousand. We quickly mocked up a black hole. What's a black hole? Well that's a collapsed star, you silly - a denormalized table with all the columns and hierarchies and attributes we need. Everything gets dumped into the denormalized table for reporting purposes, which makes it easy for a client like Jaspersoft to use as a domain. BUT.

There are foreign key inconsistencies. So when I made the black hole, I started going through a now-familiar process of iteration.  Take the facts from your clean fact table and then group them together in your first select. What we want to do is something like this:

insert into final_report

select

a.field1,
a.fact2,
a.fact3

b.attribute1,
b.attribute2,

c.label1,
c.labe22

from main_fact a
join dimension_one b on (a.key = b.key)
left outer join dimension_two c on (a.key2 = c.key)

You can see that I did just fine with my counts joining table A to table B, but table C fell short, so I performed an outer join. But that means c.label1 and c.label2 are going to be left NULL where there is no match on the a.key2. Are you with me so far? Since I know that possibility exists for every join, I may as well set up a series of errorfication fields which will let me know. That way I can left outer join all the time and not worry about my counts any more. So let's think about this in Star Trek terms. You know the scene, Spock checks the sensors and he's able to check 4 out of 5 of the facts about the alien planet, but some weird emission seems to be blocking the Enterprise' ability to get the fifth fact. With the left outer joins, you get all the facts you can get, and those that are missing can be qualified with the type of error you know ahead of time. At query time, you always have a full database for reporting and you may choose to include or exclude the partial data based on the error columns. Now generally speaking this is ugly business in a relational db, because if your 6 way join was painful in creation, imagine the time it takes to go back on it again like so:

update final_report
set missing_table_c_records = true
where label1 is null or label2 is null

Obviously the more tables you join, the more potential missing data you have. But it's better to know and present that as part of your entire analysis. Every time you update your final_report table, you have an idea of your data quality.

Enter columnar databases. In Vertica the cost of updating a single column is very low for two reasons. One, because it doesn't store the single value you change all across the database, it does a smart run length encoding. Two because it only deals with the column you select, not on 'records' as wide as your table.  The same is true for Redshift although you should tell Redshift to use the run length encoding. Either way, the columnar technologies make it very inexpensive to do field updates. And that is how we can produce full datasets that have errors built in for missing data. 

The advantages are that your end users will be able to make more sensible demands on data quality improvement. "Hey we need to get more data to populate table C', something that immediately translates into something your DBAs will understand. Furthermore, they will have a better picture of which parts of the entire data picture processed through the back end is more readily available and make appropriate judgement calls - more informed than they would be if the entire dataset were abbreviated to the net of inner joins, or blanks left with no explanation. 

Related articles

Producers, Consumers & Dynamic Data Warehousing