Thomas Kejser's Database Blog

Co-author: Lasse Nedergaard

Yesterday, Lasse ran into an issues with a query pattern in the large database that he is responsible for. Based on our conversation, we wrote up this blog and created a repro.

The troublesome query we were debugging executed like this:

1. Find a list of keys values to search for
2. Insert these keys in a temp table – lets call this the SearchFor table
3. Join the temp table to a large table (lets call it BigTable) and retrieve the full row from the large table

Why not use a correlated sub query in step 2? In this case, the customer in question had multiple code paths (including one accepting XML queries) that all needed to pass thousands of key to a final search procedure. They wanted a generic way to pass these key filters to to the final access of BigTable. The schema will make it clearer.

Schema

We created a repro and it turns out to be a great look into the strange world of database statistics. Here is the schema:

CREATE TABLE BigTable
(
  NearKey UNIQUEIDENTIFIER NOT NULL
  , Payload char(200) NOT NULL
  , RowType INT NOT NULL
)

/* Insert rows to simulate large table with 177 being the skewed value */

INSERT INTO Bigtable WITH (TABLOCK) (NearKey, PayLoad, RowType)
SELECT NEWID()
  , REPLICATE(‘X’, 200)
  , CASE WHEN n%100 BETWEEN 0 AND 50 THEN 177 ELSE n%100 END
FROM dbo.fn_nums(1000000)

/* Creates indexes and statistics */

CREATE INDEX IX_NearKey ON BigTable (NearKey)
CREATE STATISTICS STAT_RowType ON BigTable(RowType)

UPDATE STATISTICS BigTable WITH FULLSCAN

/* Insert rows that are NOT in the statistics with key = 199 */

INSERT INTO BigTable (NearKey, PayLoad, RowType)
SELECT NEWID(), ‘New Rows Not stat updated’, 199 /* New type*/
FROM fn_nums (100000)

/* Create generic search table */

CREATE TABLE #SearchFor (NearKey UNIQUEIDENTIFIER)

 

In Production, BigTable is a very large table with billions of rows. Each row in the table has RowType that describes metadata about the row. The metadata is used to populate #SearchFor before accessing BigTable, as this saves significant I/O.

Things to note about BigTable:

• All stats have been fully updated for values that are not RowType = 199. As long as we stay below this value, we have done ALL we can
• There is big skew in the RowType column.
• There are 100K values that are not in the statistics (you can run DBCC SHOW_STATISTICS (BigTable, STAT_RowType) WITH HISTOGRAM to convince yourself of this)

Query Problems

When a user queries this system, the code will first populate #SearchFor with a stored procedure like this:

Typical search procedures in the system look like this:

CREATE PROCEDURE GetDataA
AS

SELECT … INTO #SearchFor

EXEC dbo.ReturnFromBigTable

 

ReturnFromBigTable contains the troublesome statement that is the subject of this blog:

SELECT *
FROM BigTable
JOIN #SearchFor
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = <SomeValue>

 

Looks quite innocent, doesn’t it?

To simulate the customer’s workload, let us say we are searching for 1% of the rows in BigTable. We do this by populating #SearchFor.

INSERT INTO #SearchFor
SELECT TOP 10000 NearKey FROM BigTable   
TABLESAMPLE (1 PERCENT)

And now, the fun begins! Lets first use our NearKey search and for a non skewed RowType:

SELECT * FROM BigTable
JOIN #SearchFor
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 99

Plan (only join shown)

Looking at Predicate in properties for the BigTable scan we unsurprisingly see that the predicate is being driven down: [dbo].[BigTable].[RowType]=(99)… Nothing interesting so far. The plan is parallel. Again, not surprising considering that we are scanning accessing 1M rows in BigTable.

… but as Lasse and I observed on his workload, something isn’t quite right here.

Let us run the same query, but with the filter RowType = 99 replaced by RowType = 177. The new plan is:

It is at this point your design intuition should wake you up with that good ol’ WTF feeling.  “Wait a minute you should say”…before (when there were fewer rows coming out of BigTable) we were parallel and hashing the output of BigTable. Now, when there are more (50x more) rows, we are both reversing the hash order, probing into BigTable and hashing SearchFor. Note that all estimates are relatively accurate (within 10%) and statistics in the space we are searching is fully up to date… If we take a fully naïve approach, trusting relational algebra and query optimizers, there is nothing more we can do here.

But hey, it gets worse: What happens when we go outside the boundaries of the statistics. This for example, could happen when the stats are not fully updated – a completely normal situation in a database. Searching with RowType = 199 (that is not in the histogram) and estimating the plan, we get:

The estimate coming out of BigTable is 1 row, the actual is 100K; this is probably going to hurt.

Running the query gives us the deadly:

Congratulations to us, we just committed performance suicide!

Lets just list the execution times of these query variants:

Join Strategy	RowType	CPU time	Elapsed	Logical I/O Reads
Hash BigTable Probe #SearchFor	99	421ms	549ms	31429
Hash #SearchFor Probe BigTable	177	765ms	1293ms	31429+62
Merge Join (sort both)	199	1575ms	2261ms	31429+62

What is the problem here? From the user’s perspective, the query response time will often vary unpredictably – which is a poor experience. But worse, from a system and DBA perspective: the memory consumption will vary a lot depending on which query plan we end up with. This means it becomes hard to extrapolate the total system throughput and resource requirements based on the queries. If you are capitalist inclined in your mindset, let me translate this: At the end of the day, this costs you money!

Why is this happening? It is happening because there are hidden assumptions here that you have not brought into the light. We are trusting technology (in this case the database statistics and optimizer) with solving something that is a data modeling problem.

This is a very good example of something Thomas has been seeing a lot lately: An unwillingness to be specific about your requirements costs you money – a LOT of money. In other words: There is a price tag associated with ignorance. Let us make it clear what we are NOT saying: There is nothing wrong with ignorance: it is hard to know what you don’t know. Sometimes, you simply don’t know a requirement exists. The trouble arrives when denial and lack of curiosity and analysis trumps the observed data.

Fortunately, Lasse is the curious kind and he asked all the right questions and sought knowledge to remedy the situation.

Let’s get pragmatic, and talk about solutions.

The solution

With all the query plans above, which one do we want? In this case, the answer is: “None of them”

Here is an observation about the workload we discovered as we were analysing the solution: It is reasonable to assume ask the user to accept that response time is proportional with the number of results returned. From a system perspective, we also want the memory consumption and CPU to be proportional to the returned result size. These are questions you have to ask yourself as system designer and discussion you can have with users.

This is where computer science comes into the equation. We know:

1. That we can search a B-tree of n rows in lg(n) time (which is as good as constant)
2. That the the returned result are capped by the number of rows (let us call this R) in #SearchFor
3. That we are willing to accept that the runtime and resource use is proportional with number of rows returned
4. That we cant live with unpredictable behavior

The optimizer KNOWS about the search time and the cap (1+2). But the optimizer does NOT know that we are willing to declare certain properties that narrow the search spare (3) and that we want predictability (4). Our "ignorance” has a price, and the optimizer makes a choice that might be optimal, but which unfortunately is also unpredictable.

This is a great case for hinting. And we can do it like this:

SELECT *
FROM #SearchFor
INNER LOOP JOIN BigTable
  ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 177

Note something important here: The order of the joined tables matter when you use the LOOP hint (which is why we flipped them above). The above hint BOTH forces the join order and the join strategy.

An alternative, which may be more generally applicable is this rewrite:

SELECT *
FROM BigTable
INNER JOIN #SearchFor WITH (FORCESEEK)
ON BigTable.NearKey = #SearchFor.NearKey
WHERE RowType = 177
OPTION (LOOP JOIN)

The above rewrite does not guarantee join ordering. However, with the FORCESEEK hint, the plan we want is almost guaranteed.

Let us add some new rows to the result:

Join Strategy	RowType	CPU time	Elapsed	Logical I/O Reads
Hash BigTable Probe #SearchFor	99	421ms	549ms	31429
Hash #SearchFor Probe BigTable	177	765ms	1293ms	31429+62
Merge Join (sort both)	199	1575ms	2261ms	31429+62
Hinted LOOP	99	406	718	81259
Hinted LOOP	177	499	964	81259
Hinted LOOP	199	306	718	81259

Above, we can see why the optimizer didn’t want to help us! The number of IOPS required for my hinted plan are much higher than the optimizer generated plans.

The hinted plan shape is:

Obviously, we could replace the non clustered index with a clustered on (which would get rid of the RID lookup) to make this query faster. But in this case, this was not a viable solution for other reasons.

Summary

In this blog, I have shown you an example of database optimizers failing to make the “right choice”. There are several reasons this is happening:

Lack of information: The database needs declarative information about the user’s intent and what we consider the Right Choice™. In the example, the missing information is: “the speed of the returned result should always be proportional with the result size”. We used a hint to control the optimizers behavior, using our knowledge of the physical execution of queries. Other people might have taken more extreme steps and argued for noSQL – perhaps too aggressive a step in this case.

Leaky abstractions: We saw the line between the “logical” and “physical” model break down. We believed that by building the right model, we had supplied enough information to the database. The distinction between physical and logical model has, as Thomas previously argued, always been a pretty useless paradigm. We have to consider the components of the system in a holistic way, but not at the expense of maintaining the overview of individual building blocks.

Optimality vs. Good enough: the optimizers default behavior is to look for “optimal plans”. In the plan search space, even a very small space like this example, plans are sometimes found that look optimal from statistics. Yet, these plans may not be optimal or have the properties we seek – or we may fail to recognize the good plans as we search the space. In our case, we just wanted a plan that was “good enough” and had certain predictable attributes.

False hardware assumptions: The optimizer makes some assumptions about the cost of IOPS which unfortunately do not correlate with the reality of a modern machine. It assumes that optimizing for a low number of Logical IOPS is a Good Thing™. In this light, the plans generated are the right plans. However, the optimizer does not take into account that the I/O system might be fast, and that there is enough RAM to have a likelihood of finding some of the needed data already residing in the buffer pool. To prove this: Recall that we had fully updated statistics available for queries going after RowType = 177. But, running the query on a petty IBM laptop the  loop hinted plan, even on an empty buffer pool, is still faster than running the plan generated by the optimizer.

The simple query we studied here raise some interesting questions about the design tradeoffs that the data modeler and architect is forced to make. In this case, we were fortunate that Lasse was on the alert for these tradeoffs. With this example, we have seen what the potential consequences of “behavior and model ignorance” are and why you need to be alert to it. 

We are near the dawn of a new workload: BigData. While some people say that “it is always darkest just before the dawn”. I beg to differ: I think it is darkest just before it goes pitch black.  Have a cup of wakeup coffee, get your eyes adjusted to the new light, and to flying blind a bit, because the next couple of years are going to be really interesting.

In this post, I will be sharing my views on where we have been and a bit about where we are heading in the enterprise architecture space. I will say in advance that my opinions on BigData are just crystalizing, and it is most likely that I will be adjusting them and changing my mind.

Yet, I think it will be useful to go back in history and try to learn from it. Bear we me, as I reflect on the past and walk down memory lane.

The Dawn of Computing: Mainframes

If you were to look at the single most successful computing platform of ancient times (that would be before 1990), you are stuck with the choice between the Apple II, the C64 or the IBM S/360 mainframe. The first two are consumer devices, and I may get back to those in another blog post. Today, let us look at the heavy lifting server platforms, since we are after all going to talk about data.

Under the guidance of Fred Brooks, IBM created one of the most durable, highly available and performing computing platforms in the history of mankind. Today, mainframes are challenged by other custom build supercomputers, x86/x64 scale-up platforms and massive scale out systems (ex: HADOOP). But even now, the good old S/360 still holds on to a very unique value proposition. No, it is not the fact that some of these machines almost never need to reboot. It is not the prophetic beauty of JCL (a job scheduler that “gets” parallelism) or the intricacies of COBOL or PL/I…

In fact, it is not the mainframe itself that gives it an edge, it is the idea of MIPS: Paying for compute cycles!

When you pay for compute cycles, every CPU instruction you use counts. The desire to write better code goes from being a question of aesthetics, to a question about making business sense (i.e. money). Good programmers, who can reduce MIPS count, can easily visualize their value proposition to business people, and justify extraordinary consulting fees.

As we shall see, it took the rest of us a long time to realize this.

The Early 90ies: Cataclysm

My programming career really took off in the late 80ies and early 90ies – before I got bitten by the database bug. I used to write C/C++, LISP, Perl, Assembler (various), ML and even a bit of Visual Basic (sorry!) back then. Pardon my nostalgia, but in those “old days” it was expected that you were fluent in more than one programming language.

There were some common themes back then.

First of all, we took pride in killing mainframes. We saw the old green/black terminals as an early, and failed, attempt at computing – dinosaurs that had to be killed off by the cataclysm of cheap, x86 compute power (or maybe RISK processors, though I never go around to using them). We embraced new programming languages and the IDE with open arms. We thought we succeeded: IBM entered a major crisis in the 80ies for the first time in their long and proud history. However, it could be argued that it was IBM as a company that failed, not the mainframe. There are still a lot of mainframes alive today, some of them have not been turned off since the 70ies and they run a large part of what we like to call civilisation. As a computer scientist, you have to tip your hat to that.

Another theme was a general sense of quick money. IBM had a lot of fat on their organization, and all that cost had to go somewhere: It ended up as MIPS charges to the customer. This made mainframes so expensive that it was easy to compete. It was the era of the “shallow port”. ERP systems running on ISAM style “databases” were ported 1-1 to relational databases on “decentral platforms” – aka: affordable machines. Back then, this was much cheaper than running on the mainframes and it required relatively little coding effort.

The results of shallow ports was code strewn with cursors. I suspect that a lot of our hatred towards cursors is from that time. People would do incredibly silly things with cursors, because that was the easy way to port mainframe code. Oracle supported this shallow port paradigm by improving the speed of cursors and introducing the notion of sequencers and Row ID, which allows even the database ignoramus to get decent, or should we say: “does not suck completely”, performance out of a shallow port. If you hook up a profiling tool to SAP or Axapta today, you can still see the leftovers from those times.

Late 90ies: All Abstractions are Leaky

Just as proper relational data models were beginning to take off, and we began realizing the benefits of cost based optimizers and set based operations, something happened that would throw us into a dark decade of computer science. As with all such things, it started well intentioned: A brilliant movement of programmers began to worry about encapsulation and maintainability of large code bases.  Object oriented (OO) programming truly took off, it was older than that, but it now had critical mass.

The KoolAid everyone drank was that garbage collection, high level (OO) languages, IDE support and platform independence (P code/Byte Code) would massively boost productivity and that we never had to worry about low level code, storage and hardware again. To some extend, this was true: we could now trains armies of cheap programmers and throw them at them at problem (I that ironic way history plays tricks on us, the slang term was: The Mongolian Horde Technique). We also had less need to worry about ugly memory allocations and pointers – in the garbage collector we trusted. Everything was a class, or an object, and the compilers better damn well keep up with that idea (they didn’t). The champion of them all: JAVA, suffered the most under this delusion. C/C++ programmers became an endangered species. Visual Basic programmers became almost well respected. And, to please people who would give up pointers and malloc, but not curly brackets, a brilliant Dane invented C#.

At my university Danes, proud of our heritage, even invented our own languages supporting this new world (Mjolnir Beta if anyone is still interested). Everyone was high on OO. A natural consequence of this was that relational database had to keep up. If I could not squeeze a massive object graph into a relational database, that was a problem with the database, not with OO. Relational databases were not really taught at universities, it was bourgeois.

This was the dawn of object databases and the formalization of the 3-tier architecture. Central to the this architecture was the notion that scale-out was mostly about two things: 

1. Scaling out the business logic (ex: COM+, Entity Beans)
2. Scaling out the rendering logic (ex: HTML, Java Applets)

We still hadn’t figured out how to fully scale databases, though most database vendors were working on it, and there were expensive implementations that already had viable solutions (DB2 and Tandem for example). What we called “Scale-out” in the 3-tier stack was functional scale, not data scale: Divide up the app into separate parts and run those parts on different machines.

I suspect we scaled out business logic because, we believed (I did!) that this was where the heavy lifting had to be done. There was also a sense that component reuse was a good thing, and that highly abstract implementations of business logic (libraries of classes) facilitated this. Of course, we did not see the dark side: that taking a dependency on a generic component, tightly coupled you to release cycles of that component. And thus was born “DLL-hell” and an undergrowth of JAVA protocols for distributed communication (ex: CORBA, JMS, RMI)

Moving business logic to components outside the database also created a demand for distributed transactions and frameworks for handling it (ex: DTC). People started arguing that code like this was the Right Way™ of doing things:

Begin Distributed Transaction

  MyAccount = DataAccesLayer.ReadAccount()

  if  withdrawAmount <= accountBalance then

    MyAccount.Balance = accountBalance – withdrawAmount

    MyTransaction = DatabaseAccessLayer.CreateTran()

    MyTransaction.Debit = withDrawAmount

    MyTransaction.Target = MyAccount

    MyTransaction.Credit = withDrawAmount

    MyTransaction.Source = OtherAccount

    MyTransaction.Commit

 else

    Rollback Transaction

    Throw “You do not have enough money!”

 end

Commit Distributed Transaction

You get the idea… Predictably, this led to chatty interfaces. Network administrators started worrying about latency, people were buying dark fibers like there was no tomorrow. Database administrators were told to fix the problem and tune the database, which was mostly seen as a poorly performing file system. Looking back, it was a miracle that noSQL wasn’t invented back then.

Since this was the dawn of the Internet, scale out rendering made a lot of sense. It was all about moving fat HTML code over stateless protocols. It was not unreasonable to assume that you needed web farms to handle this rendering, so we naturally scaled out this part of the application. Much later, someone would invent AJAX and improve browsers to a point where this would become less of a concern.

We were high on compute power and coding OO like crazy. But like the Titanic and the final end of the Victorian technology bubble, we never saw what was about to hit us.

Y2K: Mediocre, But Ambitious

The new millennium had dawned (in 2000 or 2001, depending on how you count) and people generally survived. The mainframes didn’t break down, life went on. But a lot of programmers found themselves out of work.

In this light, it is interesting to consider that programmers were considered the most expensive part of running successful software back then. JAVA didn’t live up to its promise of cross platform delivery – it became: “Write once, debug everywhere” and people hated the plug-ins for the browser. While productivity gains from OO were clearly delivered, I personally think that IntelliSense was the most significant advance in productivity that happened between 1995 and 2005 (it takes away work from typing, so I can use it on thinking).

Professional managers, as the good glory hunters they are, quickly sniffed out the money to be made in computers during the tech bubble. As these things go, they invented a new term: “IT” that they could use to convince themselves that once you name something, you understand it. It was argued that we needed to “build computer systems like we build bridges”, but the actions we took looked more like “building computer systems like we butcher animals”. The capitalist conveyor belt metaphor, made so popular by Henry Ford and “enriched” by the Jack Welsh’ish horror regime of the 10-80-10 curve, eventually led to the predictable result: Outsourcing.

Make no mistake, I am a great fan of outsourcing. I truly believe in the world-wide, utilitarian view of giving away our technology, knowledge, money and work to poor countries. It is a great way to equalize opportunities in the world and help nations grow out of poverty – fast. In fact, I think we need to start outsourcing more to Africa now. Outsourcing it is the ultimate, global socialist principle.

The problem with outsourcing isn’t that we gave work away to unqualified people in Asia and Russia – because we didn’t – those people quickly became great, and some of the best minds in computer science were created there.  Chances are that you are reading this blog on a software written by Indians.

The problem with outsourcing was that it led to a large portion of westerners artificially inflating the demand for Feng Shui decorators, lifestyle coaches, Pilates instructors, postmodern “writers”, public servants and middle level management. These days, Europe is waking up to this problem.

But then again, if we send the jobs out of the country, all the unemployed have to do something with their time. Perhaps it is no coincidence that the computer game industry soon outgrew the movie industry. You can after all waste a lot more time playing Skyrim (amazing by the way) than watching the latest Mission Impossible, and you don’t have to support Tom Cruise’s delusions in the process

Sorry, I ramble, somebody stop me. Yet, the fact is that a lot of companies wasted an enormous amount of money sending work outside the country, dealing with cultural differences and building very poor software. One of the the results (again a big equalizer, this time of life expectancy) was that numerous people must have died from mal-practice as the doctors were busy arguing about the latest way to AVOID building a centralized, and properly implemented, Electronic Patient Journal/Record system. As far as I know, that battle is still with us.

The Rise of the East: Nothing interesting to say about XML

Once you have latched on to the idea of a fully componentized code base, it makes sense to standardize on the protocols used for interoperability. Just when we thought we couldn’t waste any more compute cycles – someone came up with SOAP and XML.

This led to a new generation of architects arguing that we need Service Oriented Architectures (SOA). There are very few of those today (both the architects and the systems), but the idea still rings true and for the customers who have adopted it. And it IS very true: a lot of things become easier if you can agree on XML as an interchange format. We also saw the rise of database agnostic data access layers (DAL). Everyone was really afraid to bind themselves to a monopoly provider.

I don’t think it is a coincidence that open standards and the rise of Asia/Eastern Europe/Russia coincided with *nix becoming very popular. The critical mass of programmers, as well as the interoperability offered by new and open standards, made it viable to build new businesses and software. The East has a cultural advantage of teaching mathematics in school at a level unknown to most westerns – I suspect it will be the end of our Western culture if we don’t adopt. Good riddance..

And the problem was again the over interpretation of a new idea. Just because XML is a great interchange format does not mean it is a good idea to store data in that format. We saw the rise (and fall) of XML database and a massive amount of programmer effort wasted on turning things that work perfectly well (SQL for example) into XML based, slower, versions of the same thing (XQuery for example). But something was true about this, and it lingered in our minds that SQL didn’t address everything, there WAS a gap: Unstructured data…

All the while, we were still high on compute power. Racking blade machines with SAN storage like there was no tomorrow. The term “server sprawl” was invented – but no one who wrote code cared. We continued to believe all that compute power was really cheap. Moore’s Law just kept on giving.

The end of the Free Lunch: Multi Core

But around 2005, it was the end of it. CPU clock speed stopped increasing. Throwing hardware at bad code was no longer an option. In fact, throwing hardware at bad code made it scale WORSE.

Those of us who had flirted with scale-out in the 90ies, and failed miserably, got flashbacks. Predictably, we acted like grumpy old men: angry and bitter that we didn’t get the blonde while she was still young and hot. Oracle became hugely successful with RAC, and the idea of good data modeling that actually has to run on hardware came back on the table, snatched from the hands of the OO fans while they were distracted by UML. People started arguing that perhaps, some of that business logic we all loved scaling out (functional scale-out, mind you), DID belong in the database after all. Thus came SQLCLR and JAVA support in Oracle and DB2. Opportunistic companies started publishing data models and selling them for good money, it was believed that if ONLY we could do 3NF modeling of data, all would be well. Inmon followers were high on data models!

People demanded automatic scale and everyone in the know, knew that this was not going to work without a fair amount of reengineering. But of course, we continued to pay homage to the idea – people don’t like their illusions broken.

While everyone was busy wrapping their head around scale-out, the infrastructure guys were beginning to show signs of panic. Big banks were complaining (and still are) that their server rooms were overflowing with machines. Their network switches were not keeping up anymore and the SAN guys went into denial: “You may measure a problem in Windows, but that is Windows who can’t do I/O, I see no problem in the SAN”. All our XML, OO and lazy code and CPU greed had started to take its toll on the environments. Somebody started talking about “Green Computing” – damn hippies!

Flashback: Kimball’s Revenge

While Teradata was busy selling “magic scale out” engines that had made great progress in this space, Kimball followers were spraying out Data Marts and delivering value in less time than it took the “EDW people” to do a vendor evaluation. The Kimball front got themselves a nice weapon of mass destruction with the Westmere and Nehalem CPU cores: cheap, powerful CPU power that took up very little space in the server room and didn’t need any special network components. It was all in-process on the same mainboard. Itanium (IA-64) went the way of the Alpha chip, the final realization that most code out there truly sucks: high clock speed beat elegant architectures.

Finally, a single machine with lots of compute power, no magic scale-out tricks, could solve a major subset of all warehouse problems (and interestingly most of the classic OLTP problems too). Vendors started digging out old 1970 storage formats that had a great fit with Kimball. Column stores and bitmap indexing got a revival. We saw the rise of Neteeza, Redbrick, Analysis Services, TM1, DB2 OLAP Views and ESSBase. This trend continues today, for example with Vertica and the Microsoft VertiPaq engines. there is even a great website dedicated to it: The BI Verdict.

Data kept growing though, and today we are seeing an interest in MPP, Kimball style, warehouses promising truly massive scale. Scale-out and -Up combined in a beautiful harmony.

But of course, social media beat us all to it. No matter how much data we tried to cram into the warehouse, the world could spam it faster than we could consume it.

BigData: More Pictures of Cats than you can Consume

Our tour of history is nearly at an end – bear with me a little longer. Today, the old dream of handling unstructured data is becoming reality – but perhaps a nightmare for some.

With BigData and HADOOP, we have a new architecture that for the first time makes some promises we can start believing:

• Storage is TRULY cheap and massive (But you have to live with SATA DAS, not SAN)
• Unstructured data, and queries on it, can be run in decent time (Not optimally fast, but good enough)
• Semi automated scale is doable (but only if you know how to write Map/Reduce or use expression trees)

This means that we can now drink the barrage of data coming from the Internet. Of course, human nature being what it is, a lot of this data is pure noise. In fact, the signal to noise ratio is scary on the Internet – I can think of no better word for it.

Among all the noise of pictures of cats, tweets and re-tweets about toilet visits, and latest news about Britney Spear’s haircut, there is something we treasure: Knowledge about human nature. The problem that BigData helps us address is that we don’t always know how to find the needle in the haystack. If we build BigData installations, we can keep the data around, unafraid of losing it or burning all our money, while we dig for that insight we so desperately seek about our customers and the world around us.

I have written elsewhere about the great fit between BigData and Data Warehouses. But I would like to reinforce my point here: Once you know the structure of the data you are looking for, by all means model it and put it in a warehouse for fast consumption. If you don’t know what you are looking for yet (or if you can wait for it), put it in BigData.

Just be careful, there is a price associated with choosing temporary ignorance over knowledge seeking.

The Cloud: Grow or Pay!

Remember those infrastructure people who complained about server rooms overflowing? Do I hear your growls of anger? As architect/programmer/DBA, I think we often have felt the frustration of dealing with infrastructure departments (in my case the SAN guys take the price). “Why can’t those people just rack the damn server and be done with it – instead of letting me wait 3 months?” we ask ourselves as we again bang our fist into the table in frustration.

But the fact of the matter is that those server rooms ARE overflowing and overheating all over the world. Handling all the complicated logistics of racking and patching servers is expensive, and that is a direct consequence of our greed for compute power. Those infrastructure people are just trying to survive the onslaught of the server sprawl – give them a break. They respond by virtualizing (SAN and virtual machines) and they fight back when we ask for hardware that is non-standard or “not certified”.

Perhaps it is time we look ourselves in the mirror and ask if we really need to waste all those CPU cycles on overly complex architectures that protect us from ourselves and our inability to properly analyse what we plan to do with data. Every time we rack a new server, power consumption goes up in the server room and it gets just a little warmer in there. We are incurring a long term cost in power and cooling by writing poor code and designing overly flexible architectures. At the end of the day, a conditional branch in the code (an “if statement”), metadata and dynamically typed, interpreted code costs CPU, and if we want to be flexible, we often need a lot of that.

If there is one thing we have learned from mainframes, it is that code lives nearly forever. Especially when it is so bad that no one dares to touch it, or so complex that no one understands it.

If it makes your day, think of this as a green initiative: you are saving the planet by writing better code and saving power. But of course, such tree hugger mentality gets us nowhere in the modern business world, so how about this instead: with the dawn (overcast?) of the cloud, compute cycles will once again be billed direct to you: the coder and DBA!

MIPS are back in a new form, and it again makes good business sense to write efficient code. In fact, it always made good sense (even in the 90ies), but all those costs you were incurring from writing bad code or forcing data into storage formats that are unnatural (e.g. XML) has so far been hidden in your corporate balance sheet. With the cloud, that cost is about to uncloak itself, and like a Klingon battleship – this might leave you in trouble!

Summary: What history Taught Us

If you are still with me here, thank you for putting up with my tantrums. It is time to wrap this up. What has history taught us about architectures and how we go about storing data?

I like to think of computing as evolutionary, not revolutionary. There are very few major breakthroughs in computer science (perhaps with the aforementioned exception of IntelliType) that don’t arrive with a lot of advance warning, and grow up slowly from many failed attempts at implementations. For example, column stores are super hot today, but they were invented back in the 70ies.

Let us have a look at the “evolutionary tree of life” for data storage engines:

Above is of course a gross over-simplification, and you could argue about where the arrows should connect. But no matter where those arrows lead, it makes the point I want to make: That nearly all of these technologies are still around today and very much alive (perhaps with the exception of XML databases, but those were pretty silly to begin with). In fact, most large organisations have most of them around.

Are all these technologies just alive because they have not yet been replaced with the latest and greatest? Is noSQL, HADOOP or some other new product bound to replace them all over time?

I think the different technologies are there for a good reason. A reason very similar to why there are many species of animals in the world. Each technology is evolved to solve a certain problem, and to live well in a specific environment. When you target the right technology to a certain problem (for example column stores towards dimensional models) you solve that problem elegantly and with the least amount of compute cycles – preserving energy. Very soon, solving problems with a low number of compute cycles in elegant ways is going to really count.

While you might be able to solve most of the structured DW data problems with HADOOP, you eventually have to ask: Is a generic map/reduce paradigm really the way to go about that?

Humans have a strange way of wanting a single answer to the big and complex question in life, and we waste significant time searching for it. Entire cults of management theories (and I use that term lightly) and religions are build around these over simplifications. We seek simple answers to complex question, and as IT architects, this can lead us down a path of believing in “killer architectures”.

What history has taught us is that such killer architectures do not exist, but that the optimal strategy (just like in nature) depends on which environment you find yourself in. Being an IT architect means rising above all these technologies, to see that big picture, and resisting the temptation to find a single solution/platform for all problems in the enterprise. They key is to understand the environment you live in, and design for it using a palette of technologies.

Staying in the metaphor, this leads me to another conclusion: Just like evolution made us store fat on our bodies in different places depending on the usage (thank goodness) you also have to consider the option of storing your data in more than one place. Having a bit of padding all over the body is a lot more charming (and healthy) than a beer belly

References:

• The Data Modeling Category on this site with more ramblings
• Fred Brooks wrote a classic, “The Mythical Man Month”, about his experiences working on the S/360
• Ralph Kimball’s books is all you need to read to get started on data warehousing
• Louis Gerstner, former CEO of IBM, wrote a great book about the fall and rebirth of this great company: “Who Says Elephants Can’t Dance”

Hash distributing rows is a wonderful trick that I often apply. It forms one of the foundations for most scale-out architectures. It is therefore natural to ask which hash functions are most efficient, so we may chose intelligently between them.

In this blog post, I will benchmark the build in function in SQL Server. I will focus on answering two questions:

• How fast is the hash function?
• How well does the hash function spread data over a 32-bit integer space

I know there is also the question about how cryptographically safe the function is, but this is not a necessary property for scale-out purposes – hence, that aspect is out of scope for this blog.

For the picky people (my dear soul mates): I will be using 1K = 1024 in the following, so 64K = 65536.

Test data

Because we sometimes find ourselves hashing dimension keys, I will use a very simply table with a nice key range as my test input. Such a key range could for example have been generated by an IDENTITY(1,1) column or a SEQUENCER. Here is the test script that generates the input:

CREATE TABLE CustKey (SK INT NOT NULL)

INSERT INTO CustKey WITH (TABLOCK) (SK)
SELECT n FROM dbo.fn_nums(10000000)

 

(see my previous post for the fn_nums source)

Speed of the Hash function

SQL Server exposes a series of hash functions that can be used to generate a hash based on one or more columns.

The most basic functions are CHECKSUM and BINARY_CHECKSUM. These two functions each take a column as input and outputs a 32-bit integer.

Inside SQL Server, you will also find the HASHBYTES function. This little gem can generate hashes using MD2, MD4, MD5, SHA and SHA1 algorithms. The problem for the purpose of our test is that these function spit out BINARY types, either 128 bits (for SHA) or 160 bits (for MD). However, we can quickly map that value into the integer space by doing a modulo (in SQL: %) with MaxInt (2**31 – 1). Interetingly, Modulo work directly on values of BINARY in SQL Server.

Before I move on the results, I would like share a little trick. When you need to get the raw CPU speed of the hash, you may be tempted to do something like this:

SELECT BINARY_CHECKSUM(SK) AS h
FROM CustKey

 

But this is wrong. Why? Because you are both measuring the time it takes to hash as well as the time it takes to transfer the rows to the client. This is not very useful, on my laptop it takes over two minutes to run the above query.

Instead, you should try to measure the raw time the server needs to calculate the data. One way to achieve this (and this trick can be used on most queries you benchmark) is to wrap the SELECT statement in a MAX function. Like this:

SELECT MAX(h)
FROM (
    SELECT BINARY_CHECKSUM(SK) AS h
    FROM CustKey
    ) noRows
OPTION (MAXDOP 1)

Notice something else interesting, I am forcing parallelism down to 1. I want to get as close to the raw cost of the CPU as possible – so I don’t want to clutter the query with any overhead of parallelism. The query now runs in a little less than 3 seconds on my laptop.

I addition to the above trick, we also need to quantify the time SQL Server spends on just accessing the table. We can get a good approximation of this by measuring the SELECT statement without any hash function. On my laptop, I measured this to be around 3200 ms. In the results below, I have subtracted this runtime so we are only measuring the cost of the hashing.

With all this said, let us have a look at the results:

Some comments on the data:

• Notice that MD2 is very inefficient. I have not studied this algorithm in great detail and would appreciate a specialist commenting here.
• Apart from the MD2 anomaly, the HASHBYTES functions all performan pretty much the same. Using around 1 micro sec per hash value calculation
• BINARY_CHECKSUM and CHECKSUM are MUCH faster, by about an order of magnitude
• Doing a simple modulo that spreads out the data over the integer space is about the same speed as CHECKSUM and BINARY_CHECKSUM
• It is not the cast to NVARCHAR that uses CPU time (HASHBYTES requires NVARCHAR input)
• However, the checksum function do take longer to run when you first cast to NVARCHAR. Costing around 100 ns per INT value

Of course, the HASHBYTES functions also have other characteristics than speed which may be desired (cryptographic utility for example). but if you you want is speed, it looks like CHECKSUM and BINARY_CHECKSUM are faster.

But, let us see how good the functions are at spreading the data.

Spread of the Hash Function

Another desirable characteristic, apart from speed, of a hash function is how well it spreads values over the target bit space. This is useful not only for cryptographic purposes, but also to “bucket” data values into equal sized portions, for example in scale-out MPP architectures.

Often, you will not know in advance how many buckets you eventually want to subdivide the integer space into. You may start out with 32K buckets and divide up the integer interval in 128K (32-bit space divided by 128K = 32K) even sized buckets, each with 128K values in them:

• Bucket0: [MinInt…MinInt + 128K[
• Bucket1: [MinInt + 128K…MinInt + 128K*2[
• Bucket2: [MinInt + 128K*2…MinInt + 128K*3[
• … etc..
• Bucket 32K: [MinInt + 128K*32K…MaxInt]

Perhaps you will later want a further subdivision into 64K buckets, each with 64K hash values.

Wise from the runtimes I measured before, I used only 1M rows with values [1…1M] for this test. However, this is plenty to show some good results, so don’t worry.

To avoid running chi-square testing (See later) of integer value spread all over the 32-bit space, I calculated the hash of each key and I then bucketed them into 64K bucket ranges (each with 64K values, neat isn’t it?).

This table comes in handy for that bucketing

CREATE TABLE ExpectedBuckets 
(
    BucketID INT
  , RangeStart INT
  , RangeEnd INT)

INSERT INTO ExpectedBuckets
SELECT
    n
    ,(n – 1 – POWER(2,15)) * POWER(2,16)
    , (n – POWER(2,15)) * POWER(2,16) – 1
FROM dbo.fn_nums( POWER(2,16) )

A sample output form this table:

I can calculate the hash values and put them into a table like this:

CREATE TABLE HashResult
(
  ObservationID INT IDENTITY(1,1)
, HashValue_i INT
)

CREATE CLUSTERED INDEX CIX ON HashResult (HashValue_i)

 

And finally, I can join them all up and count the values in each bucket:

SELECT
      BucketID
    , CAST(ISNULL(COUNT(HashValue_i), 0) AS FLOAT) AS Observed
FROM ExpectedBuckets EB
LEFT JOIN HashResult
ON HashValue_i BETWEEN EB.RangeStart AND EB.RangeEnd
GROUP BY BucketID

 

We now have a nice count of the content in hash bucket, it looks like this:

How do we test if this is a good result or not? It seems that min/max, averages and standard deviations don’t really suffice here (sorry modern MBA types, I realise I am going to loose you now ).

We need to do some hypothesis testing. The hypothesis I want to test (aka: the NULL hypothesis – all resemblance to relational algebra is purely coincidental).

“My chosen hash function evenly distributes the integers into the 64K hash buckets”

Interestingly, this hypothesis makes a prediction (has to, if not we cannot falsify it): namely that in each bucket filled by hashing the 1M rows, we should expect to measure 1M / 64K = 15 members.

There is also an important twist here: We have to measure the content of each bucket, even when it is zero. This is why I use a left join in the query above, we have to measure all the results we know, not just count the hash buckets that actually get filled (you can see how not bucketing the data would have made the test result very large).

We can now test the hypothesis against the measured result. If we see correlation, then it gives us confidence in the correctness of the hypothesis, and lets us conclude that our hash function is a good one. It also gives us the ability to compare different hash functions with each other. We can test for correlation between the predicted result and the actual result with the chi-square test. Interestingly, the tools we need for this test is even exposed in Excel as the CHISQ series of functions.

Without further ado (and insults), here are the results:

(source data available on request)

Some observations about the data:

• Most importantly: the CHECKSUM and BINARY_CHECKSUM functions are very poor hash functions. Even with 1M input and 64K large buckets, they don’t even fill the hash space. The skew is massive on non-skewed input data.
• We also see that casting the INT to NVARCHAR improves the behavior of both BINARY_CHECKSUM and CHECKSUM, but not enough to make them good hash functions.
• SHA and SHA1 seem to give us the best spread of the data, but at 1 microsecond per hash, it seems rather expensive.
• Notice that the MD series do not give a particularly nice spread. I suspect this is related to the way I divide the data with MaxInt to get from the 120bit value to a 4-byte integer
• We see that only the modulo does a distribution that gives us nearly 100% confidence (not fully though, because 1M does not divide 64K, which is reflected in the 824 value of Chi-squared). This result is expected, and my function here is engineered to “game the test”. The modulo cannot generally be used like this to spread the value all over the 32 bit integer space.
• For fun, I put in a test where I use random values all over the integer space. Even with a confidence interval of 10%, we cannot say it distributes evenly. This means that SHA functions distribute better than random on this input dataset. Your mileage will of course vary for random values

On the note of the CHECKUM series, Books Online does (under?) state:

If one of the values in the expression list changes, the checksum of the list also generally changes. However, there is a small chance that the checksum will not change. For this reason, we do not recommend using CHECKSUM to detect whether values have changed, unless your application can tolerate occasionally missing a change. Consider using HashBytes instead. When an MD5 hash algorithm is specified, the probability of HashBytes returning the same result for two different inputs is much lower than that of CHECKSUM.

Summary

In this blog I have explored properties of hash function built into SQL Server. The results indicate some general guidance that could be followed:

• For best spread over the hash space, use SHA or SHA1 algorithms with HASHBYTES function
• Be VERY careful about using CHECKSUM and BINARY_CHECKSUM as these functions do not spread a hashed integer value nicely over the full space. Depending on what you use these functions for, they may not be suitable
• It takes about 1 CPU microsecond to calculate a single call to HASHBYTES, except when hashing with MD2
• Using HASHBYTES to spread out values over the 32-bit integer space seems rather expensive, since even the best cases uses tens of thousands of CPU instructions.
• This indicates that it may be possible to implement a better hash function if you do not care about cryptographic properties of the hash

Reference:

• Grade of the Steel Category

History tracking in warehouses is a controversial discipline. I this post, I will begin to unravel some of the apparent complexities by taking apart the history tracking problem, piece by piece.

Observation: History tracking an EDW is not the same as archiving the source.

I already blogged about using HADOOP or "dumb databases" for archival or source system data. The purpose of these archives is to avoid the discussion with the business user about loosing data. We can always bring back the source. I also voiced the opinion that you should not bother with modeling the source, and that it is trivial to automate this archival process in such a way that the source can be restored to any point in time.

Observation: history tracking is part of the business requirements for a warehouse.

We will only track history on data that the end user needs history about. We will not model history just for the sake of modeling it.

Observation: Fact table history is very different from dimension table history

I hope it is clear that tables containing fact data typically don’t act like dimension/reference tables. I will assert that OLTP designers (they guys we get our source data from) have known and agreed upon this simple observation for years: tables behave differently and fall into certain patterns. For example: Most OLTP system have tables that are "add only", which means we don’t need to track history on them (they are already historized). As another example: By following the well laid out arguments in Kimball books on how to models inventory systems, it is also clear that we need a modeling trick "snapshots" which is quite distinct from they way we model dimensions,

In other words: Just because a fact table has from/to date does not mean we used a type 2 dimension pattern to model it.

Dimensions, on the other hand, have a property that facts rarely have: The history we model on these tables depends a lot on what we are trying to achieve for the business user and their behavior is not dictated by the behavior of the source.

Modeling Facts

Kimball has written extensively about modeling fact data, and I mostly agree with his observations and techniques. I am not even sure this statement is disagreement, but: I would like to point out that all the talk about "choosing the grain" generally doesn’t make any sense to me: just pick the lowest possible grain you can get from the source.

Modeling Dimensions

After this initial tap-dance, let me turn my attention to the subject I really want to talk about: modeling dimensions and the problem with surrogate keys.

Let me start from a set of commonly perceived problems for Kimball style warehouses. Imagine you have a model like the below:

Now extend you mental image to include a situation where we change our mind about the history tracking in the customer dimension. We obtain knowledge (maybe through data cleansing) about the education of our customers. There are two ways to handle this:

1) In the simplest model, we may simple add this information "going forward". Our old, historical records remain unchanged, but we can analyze the future data by education. This is not the problem I am concerned about, and I think this modeling trick has limited use.

2) In the more complex model, we may wish to change history "backwards in time". This means that the data model may change like this:

It is the complex model that seems to worry the opponents of dimensional modeling for Enterprise Warehousing, especially the transformation indicated by the orange arrow above.

Here is the crux of the argument against dimensional modelers, as I have understood it from my many discussion with warehouse specialists:

Problem 1: changing your mind about history means that ALL fact tables that reference the Customer dimension must change to reference the new keys in Customer’. For some reason, this is perceived as a large dependency problem that is insurmountable at scale.

Answer 1: if you don’t have a meta model (or naming convention) that allow you to quickly (literally in minutes) analyse the impact on your database of making such a change, you are doing something very wrong with your dimensional model. This problem has been solved several times by builders of dimensional enterprise warehouses – I am sure you can solve it too.

Problem 2: changing the history tracking of the dimension requires large update statements to all facts that reference it.

Answer 2: this issue has some merit, but it depends on the implementation of the database you work on. It is true that in a multi TB system, such an update can be painful. But even in an SMP system, I would temper this statement with the comment: "but not as painful as you may think". This will be the subject of a future post.

We should also observe that the impact on fact tables means that the surrogate keys we use in Kimball models, are not good keys (as per my previous definition). The reason they are not good keys, is that Kimball surrogate keys track two things:

1. The uniqueness of the dimension keys
2. The history of the entity

When one of these two are volatile (the history) the entire model becomes volatile. This is not a desirable trait of a large warehouse.

Problem 3: not every business user wants the same version of history tracking, so a single type2 dimension table is wrong.

Answer 3: this is also true for some scenarios, and it hints at a model where there are multiple versions of each dimension and a new way to link facts and dimensions together. We observe that several warehouses are built where this never happens, so we should be careful not to over-generalize the model technique to always take an edge case like this into account.

Here we observe that this additional requirement does not mean that we need to abandon the notion of type 1 columns – we just have an additional requirement which we could call "multiple views of type 2 columns".

Summary

The three above problems (of which only the last two have any merit) are quoted, in different forms, by the persons I have heard argue that dimensional models are unsuitable for an "Enterprise" Data Warehouse.

The solution, and this baffles me, seems to be: use a 3NF data model instead of Kimball’s model. I think this 3NF approach is a fallacy and I have witnessed the horrors that arise from it. But as I have admitted, the arguments DO have merit and should be addressed in the model.

Location:Boing 767 over the Pacific

References:

• Data Warehouse Category
• Modeling Category
• Good keys, what are they like?
• Data Warehouse reference architecture – Staging and Archive (external link – great related post by Marcel Franke)

I am happy to announce that companion volumes for performance tuners and operations people are now again, for the first time since Analysis Services 2000, a reality.

As a developer, you can learn about building, tuning and troubleshooting Analysis Services 2005, 2008 and 2008R2 cubes(yes, the guides cover all three editions) in the Performance Guide.

If you are the DBA or operations team, you can read about running such cubes in production in the Operations Guide.

And finally, if you are a consultant, expert developer or cube DBA, you can learn how to build the meanest and largest cubes from the 5 day Analysis Services Maestro Course.

These three artifacts, which I am proud to have contributed to (and for the guides, leading the effort on), are a big milestone in my Analysis Services career. They represent a large amount of knowledge transfer from Microsoft to the field. The publication of the companion volumes also marks my transition into some new and exiting projects at least for the near future. I will busy be digging into more “grade of the steel” work,  among it ROLAP UDM testing, and I hope to blog about over at SQLCAT.com very soon.

Thanks to everyone for the incredible feedback during the writing of these guys. The Analysis Services community is very vibrant and these guides are the result of our collaboration in the field, and I am happy to give something back.

References:

• Analysis Services 2008R2 Performance Guide
• Analysis Services 2008R2 Operations Guide

Throughout history, similar ideas tend to surface at about the same time. Last week, at SQLBits 9, I did some “on stage” tuning of the Paul Randal INSERT challenge.

It turns out that at almost the same time, a lab run was being done that demonstrated, on a real world workload, a technique similar to the one I ended up using. You can find it at this excellent blog: Rick’s SQL Server Blog.

Now, to remind you of the Paul Randal challenge, it consists of doing as many INSERT statements as possible into a table of this format (the test does 160M inserts total)

CREATE TABLE MyBigTable (
    c1 UNIQUEIDENTIFIER ROWGUIDCOL DEFAULT NEWID ()
    ,c2 DATETIME DEFAULT GETDATE ()
    ,c3 CHAR (111) DEFAULT ‘a’
    ,c4 INT DEFAULT 1
    ,c5 INT DEFAULT 2
    ,c6 BIGINT DEFAULT 42);

Last week, I was able to achieve  750K rows/sec (runtime: 213 seconds) on a SuperMicro, AMD 48 Core machine with 4 Fusion-io cards with this test fully tuned. I used 48 data files for best throughput, the subject of a future blog.

Now, I suspect Paul used NEWID() for key generation, to create as much I/O as possible with page splits. If the goal is purely to optimize throughput, why not ask: Is NEWID really the best way to generate a key?

Traditional SQL Server knowledge dictates that either IDENTITY(1,1) or NEWSEQUENTIALID are better key choices: they lay out the keys sequentially, at the end of the index. But one should not easily trust such recommendations without validating them. I did the above test on the 48 core box Fusion-io lend me, and here are the results:

 

This might shock you! The “Traditional, sequential layout” of the key is more than 100 times SLOWER than completely random inserts!

The CPU are NOT at 100% during this (they were when using NEWGUID). The problem shows itself when we dig into sys.dm_os_wait_stats. (by the way, below is my new: “this is how I think about it” diagram, let me know if you like this way of doing things and I will continue to use this format for illustrating troubleshooting mind processes)

We are waiting for a PAGELATCH on the last pages of the index. This wait dominates the run. The price of page splits is minor (when you have a fast I/O system) compared to the price of coordinating the last page memory access!

As Rick writes in his blog, there are ways to work around this problem by being smart about your key generation. Basically, you want a key that is NOT sequential. Rick flips the bits around in a Denali SEQUENCE objects – a very neat idea that uses a new SQL Server Denali feature.

For my test at SQLBits (which ran on 2008R2), I used a variant of the same idea: I took the @@SPID of the users session and added an increasing value (increments controlled by the app server) to each insert made. On reflecting, using the SPID is not a good idea, you should probably generate the key something like this:

INSERT Key =

   [Unique App Server ID] * 10**9 + [Unique Number Generated at App Server]

I also tried the hash trick I have described here – to compare the techniques.

Below are the results of my test of Paul’s INSERT Challenge with different key strategies (I have taken out the sequential ones, since we have seen how poorly they perform)

The last column is a trick I will describe in more detail in a later blog. Suffice to say that I reached 1.1M INSERT/sec (runtime: 150 sec) using a key generation strategy very similar to Rick’s.

Edits:

• The UNIT in the introduction was originally wrong. it should of course be 750K not 750M as can be seen later in post.

Lately, I have been discussing data modeling with a lot of people and there is simply so much ground to cover. Lots of exciting developments – it seems I was on to something when I started a blog about this topic.

To make faster progress, I have teamed up with Marcel Franke and  Cardory Van Rij.  I am very exciting about this development, We will be doing some blogging together in the near future, both here on blog.kejser.org and on their individual blogs.

Based on feedback on my previous post, I will be writing about Type2 dimensions and history tracking for the next post in the modeling series. It will be a little delayed, since I am traveling a lot the next month.

Stay tuned…

I know some of you track Conor Cunningham’s blog, which I highly recommend. The information on query optimizers is rather sparse out there – and it is a real privilege when people who know what they are talking about share their information freely on a blog. There are just not that many of them.

Here is something to think about in that context: Even if you are only interested in SQL Server, there are good things to be learned from studying other database engines – they are after all very similar. One of the good blogs for that purpose is the Oracle optimizer blog: http://blogs.oracle.com/optimizer/. It is delightfully free from marketing, and packed with good stuff. Don’t worry about the fact that they like textual query plans over there in Oracle. You might be used to viewing plans graphically, but I am sure you agree that such minor difference in display preferences can be set aside in the name of reading good material.

Now that the cat is out of the bag on the Kimball forum, I figured it would be a good idea to present the full architecture that I will be arguing for. I was hoping to build up to it slowly, by establishing each premise on its own before moving on to the conclusion.

But perhaps it is better to start from the conclusion and then work my way down to the premises and show each one in turn.

Here is the architecture and data flow I will be arguing should form the basis of a healthy, performing, flexible and solid foundation for an enterprise:

 

Method, at a High Level

The basic method for building this is:

• Model your data as Kimball would have done it. But ONLY model Type1 history in the dimension. The dimension form the entities (but not the entity history)
• If you believe snowflakes are easier to maintain, go ahead and normalize the entities, but do so using standard OLTP methods (you are allowed to have 1-m keys)
• Use the maps as described earlier to maintain key invariance
• For the attributes you want to history track, add a entity history table and add only the attributes you want to track (if not, you get into some interesting ETL issues later)
• If users want fact tables with type2 history, either create a mart for them (using the join technique I will describe), materialize a view, or allow them directly on the EDW using joins that a slight more complicated than a traditional star schema.
• Users who only need type1 history can simply use the EDW directly

Major choices

There are some major Kimball choices that must be made to make this model work

• The grain is always the lowest possible level. This means that the facts in the EDW (though maybe not in the mart, if you use it) are copies of the source transactions. This gets rid of any complaints about aggregating the fact and losing data
• The world is dimensional, and we are letting users lose on the EDW itself
• We allow either normalized or pure star schema dimensions/entities. Our choice in this will depend on our target database engine and required performance.
• We are using maps to keep the entity keys stable. This minimized the need for reload of facts
• We are assuming that it is easy to reload entities/dimensions and change their structure (as long as the key remains stable). If you are having trouble handling a billion row customer table, ask questions on this blog as we move forward.
• Facts contain all the data of the source. If there is “garbage” in the source we will split that out into a vertically partitioned table off the fact and put it on slow (i.e. cheap) storage.
• Facts are fully de-normalized – there are no big table/big table joins here

What happened to my ODS/Vault/Staging/Archive etc?

One thing to notice is what is NOT in this picture: There is no ODS, Vault or any 3NF databases. The reason is that I am of the opinion that you don’t need one, at least not one that has a data model you care about. I previously said that I would make the point that you should not bother modeling the source, it is a waste of time. This is an argument that needs a lot more backing, more about this later. But a few observations:

I acknowledge that there are auditing requirements which may drive you towards a “zero information loss” model, and that the EDW is often tasked with solving this problem. Note that zero information loss is achievable in the model above, with a few tweaks to the Kimball model. Here is how:

One of the things we can leverage today is the new HADOOP type storage systems (like HIVE or Cassandra). These systems finally deliver on the promise of truly cheap storage – no big SAN storage racks involved: cheap hardware all the way down to the disks. They can run massive archives on SATA spindles, maintain multiple copies of the data and run ad-hoc analysis with high block size I/O (giving you reasonable speed, even on SATA) at record low cost/TB. They are not (yet) super fast at responding to ad-hoc queries at high user concurrency, typically taking many minutes or hours to respond. But responding to ad-hoc queries fast and providing the “current business view of the truth” is what the EDW is all about.

In this architecture we use the HADOOP both as storage for raw copies of the source system and as a source itself (for example for web log data). HADOOP is the mother of all stovepipes and it lets us bring back old data if required to do so, satisfying the auditors and de-railing the argument about keeping too much data around just for this purpose. In other words: it leaves us free to model the EDW to only contain data that we know is needed to answer business problems and throw the rest away. We don’t want to understand the source or model it (except for the purpose of ETL’ing from it for the exact structures we need). We just make sure we don’t lose the source, if it cannot guarantee zero loss for us.

If we have them,  analysts are also free to hack away with ad-hoc queries on raw source data in HADOOP, in the table format and model they are used to see in the source, without having to disturb the source or rely on some new vocabulary or over normalized model technique.

Moving forward (or back) from here

I am fully aware that I make some grand claims, and of course, you should not believe me until you have seen the arguments, the design considerations, the safety measures, the methods, the raw numbers and calculations.

But this is a start for the debate and a framework for fleshing out these arguments.

References:

• Data Warehouse Introduction