Should OLAP databases be denormalized for read performance? [closed]

Closed. This question is opinion-based. It is not currently accepting answers.

---

Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.

Closed 2 years ago.

Improve this question

I always thought that databases should be deno开发者_如何学运维rmalized for read performance, as it is done for OLAP database design, and not exaggerated much further 3NF for OLTP design.

PerformanceDBA in various posts, for ex., in Performance of different aproaches to time-based data defends the paradigm that database should be always well-designed by normalization to 5NF and 6NF (Normal Form).

Have I understood it correctly (and what had I understood correctly)?

What's wrong with the traditional denormalization approach/paradigm design of OLAP databases (below 3NF) and the advice that 3NF is enough for most practical cases of OLTP databases?

For example:

• "The simple truth ... is that 6NF, executed properly, is the data warehouse" (PerformanceDBA)

I should confess that I could never grasp the theories that denormalization facilitates read performance. Can anybody give me references with good logical explanations of this and of the contrary beliefs?

What are sources to which I can refer when trying to convince my stakeholders that OLAP/Data Warehousing databases should be normalized?

To improve visibility I copied here from comments:

"It would be nice if participants would add (disclose) how many real-life (no science projects included) data-warehouse implementations in 6NF they have seen or participated in. Kind of a quick-pool. Me = 0." – Damir Sudarevic

Wikipedia's Data Warehouse article tells:

"The normalized approach [vs. dimensional one by Ralph Kimball], also called the 3NF model (Third Normal Form) whose supporters are referred to as “Inmonites”, believe in Bill Inmon's approach in which it is stated that the data warehouse should be modeled using an E-R model/normalized model."

It looks like the normalized data warehousing approach (by Bill Inmon) is perceived as not exceeding 3NF (?)

I just want to understand what is the origin of the myth (or ubiquitous axiomatic belief) that data warehousing/OLAP is synonym of denormalization?

Damir Sudarevic answered that they are well-paved approach. Let me return to the question: Why is denormalization believed to facilitate reading?

---

Mythology

I always thought that databases should be denormalized for reading, as it is done for OLAP database design, and not exaggerated much further 3NF for OLTP design.

There's a myth to that effect. In the Relational Database context, I have re-implemented six very large so-called "de-normalised" "databases"; and executed over eighty assignments correcting problems on others, simply by Normalising them, applying Standards and engineering principles. I have never seen any evidence for the myth. Only people repeating the mantra as if it were some sort of magical prayer.

Normalisation vs Un-normalised

("De-normalisation" is a fraudulent term I refuse to use it.)

This is a scientific industry (at least the bit that delivers software that does not break; that put people on the Moon; that runs banking systems; etc). It is governed by the laws of physics, not magic. Computers and software are all finite, tangible, physical objects that are subject to the laws of physics. According to the secondary and tertiary education I received:

• it is not possible for a bigger, fatter, less organised object to perform better than a smaller, thinner, more organised object.

• Normalisation yields more tables, yes, but each table is much smaller. And even though there are more tables, there are in fact (a) fewer joins and (b) the joins are faster because the sets are smaller. Fewer Indices are required overall, because each smaller table needs fewer indices. Normalised tables also yield much shorter row sizes.

• for any given set of resources, Normalised tables:

  • fit more rows into the same page size
  • therefore fit more rows into the same cache space, therefore overall throughput is increased)
  • therefore fit more rows into the same disk space, therefore the no of I/Os is reduced; and when I/O is called for, each I/O is more efficient.
    .
• it is not possible for an object that is heavily duplicated to perform better than an object that is stored as a single version of the truth. Eg. when I removed the 5 x duplication at the table and column level, all the transactions were reduced in size; the locking reduced; the Update Anomalies disappeared. That substantially reduced contention and therefore increased concurrent use.

The overall result was therefore much, much higher performance.

In my experience, which is delivering both OLTP and OLAP from the same database, there has never been a need to "de-normalise" my Normalised structures, to obtain higher speed for read-only (OLAP) queries. That is a myth as well.

• No, the "de-normalisation" requested by others reduced speed, and it was eliminated. No surprise to me, but again, the requesters were surprised.

Many books have been written by people, selling the myth. It needs to be recognised that these are non-technical people; since they are selling magic, the magic they sell has no scientific basis, and they conveniently avoid the laws of physics in their sales pitch.

(For anyone who wishes to dispute the above physical science, merely repeating the mantra will no have any effect, please supply specific evidence supporting the mantra.)

Why is the Myth Prevalent ?

Well, first, it is not prevalent among scientific types, who do not seek ways of overcoming the laws of physics.

From my experience, I have identified three major reasons for the prevalence:

1. For those people who cannot Normalise their data, it is a convenient justification for not doing so. They can refer to the magic book and without any evidence for the magic, they can reverently say "see a famous writer validates what I have done". Not Done, most accurately.

2. Many SQL coders can write only simple, single-level SQL. Normalised structures require a bit of SQL capability. If they do not have that; if they cannot produce SELECTs without using temporary tables; if they cannot write Sub-queries, they will be psychologically glued to the hip to flat files (which is what "de-normalised" structures are), which they can process.

3. People love to read books, and to discuss theories. Without experience. Especially re magic. It is a tonic, a substitute for actual experience. Anyone who has actually Normalised a database correctly has never stated that "de-normalised is faster than normalised". To anyone stating the mantra, I simply say "show me the evidence", and they have never produced any. So the reality is, people repeat the mythology for these reasons, without any experience of Normalisation. We are herd animals, and the unknown is one of our biggest fears.

   That is why I always include "advanced" SQL and mentoring on any project.

My Answer

This Answer is going to be ridiculously long if I answer every part of your question or if I respond to the incorrect elements in some of the other answers. Eg. the above has answered just one item. Therefore I will answer your question in total without addressing the specific components, and take a different approach. I will deal only in the science related to your question, that I am qualified in, and very experienced with.

Let me present the science to you in manageable segments.

The typical model of the six large scale full implementation assignments.

• These were the closed "databases" commonly found in small firms, and the organisations were large banks
• very nice for a first generation, get-the-app-running mindset, but a complete failure in terms of performance, integrity and quality
• they were designed for each app, separately
• reporting was not possible, they could only report via each app
• since "de-normalised" is a myth, the accurate technical definition is, they were un-normalised
  • In order to "de-normalise" one must Normalise first; then reverse the process a little in every instance where people showed me their "de-normalised" data models, the simple fact was, they had not Normalised at all; so "de-normalisation" was not possible; it was simply un-normalised
• since they did not have much Relational technology, or the structures and control of Databases, but they were passed off as "databases", I have placed those words in quotation marks
• as is scientifically guaranteed for un-normalised structures, they suffered multiple versions of the truth (data duplication) and therefore high contention and low concurrency, within each of them
• they had an additional problem of data duplication across the "databases"
• the organisation was trying to keep all those duplicates synchronised, so they implemented replication; which of course meant an additional server; ETL and synching scripts to be developed; and maintained; etc
• needless to say, the synching was never quite enough and they were forever changing it
• with all that contention and low throughput, it was no problem at all justifying a separate server for each "database". It did not help much.

So we contemplated the laws of physics, and we applied a little science.

We implemented the Standard concept that the data belongs to the corporation (not the departments) and the corporation wanted one version of the truth. The Database was pure Relational, Normalised to 5NF. Pure Open Architecture, so that any app or report tool could access it. All transactions in stored procs (as opposed to uncontrolled strings of SQL all over the network). The same developers for each app coded the new apps, after our "advanced" education.

Evidently the science worked. Well, it wasn't my private science or magic, it was ordinary engineering and the laws of physics. All of it ran on one database server platform; two pairs (production & DR) of servers were decommissioned and given to another department. The 5 "databases" totalling 720GB were Normalised into one Database totalling 450GB. About 700 tables (many duplicates and duplicated columns) were normalised into 500 unduplicated tables. It performed much faster, as in 10 times faster overall, and more than 100 times faster in some functions. That did not surprise me, because that was my intention, and the science predicted it, but it surprised the people with the mantra.

More Normalisation

Well, having had success with Normalisation in every project, and confidence with the science involved, it has been a natural progression to Normalise more, not less. In the old days 3NF was good enough, and later NFs were not yet identified. In the last 20 years, I have only delivered databases that had zero update anomalies, so it turns out by todays definitions of NFs, I have always delivered 5NF.

Likewise, 5NF is great but it has its limitations. Eg. Pivoting large tables (not small result sets as per the MS PIVOT Extension) was slow. So I (and others) developed a way of providing Normalised tables such that Pivoting was (a) easy and (b) very fast. It turns out, now that 6NF has been defined, that those tables are 6NF.

Since I provide OLAP and OLTP from the same database, I have found that, consistent with the science, the more Normalised the structures are:

• the faster they perform

• and they can be used in more ways (eg Pivots)

So yes, I have consistent and unvarying experience, that not only is Normalised much, much faster than un-normalised or "de-normalised"; more Normalised is even faster than less normalised.

One sign of success is growth in functionality (the sign of failure is growth in size without growth in functionality). Which meant they immediately asked us for more reporting functionality, which meant we Normalised even more, and provided more of those specialised tables (which turned out years later, to be 6NF).

Progressing on that theme. I was always a Database specialist, not a data warehouse specialist, so my first few projects with warehouses were not full-blown implementations, but rather, they were substantial performance tuning assignments. They were in my ambit, on products that I specialised in.

Let's not worry about the exact level of normalisation, etc, because we are looking at the typical case. We can take it as given that the OLTP database was reasonably normalised, but not capable of OLAP, and the organisation had purchased a completely separate OLAP platform, hardware; invested in developing and maintaining masses of ETL code; etc. And following implementation then spent half their life managing the duplicates they had created. Here the book writers and vendors need to be blamed, for the massive waste of hardware and separate platform software licences they cause organisations to purchase.

• If you have not observed it yet, I would ask you to notice the similarities between the Typical First Generation "database" and the Typical Data Warehouse

Meanwhile, back at the farm (the 5NF Databases above) we just kept adding more and more OLAP functionality. Sure the app functionality grew, but that was little, the business had not changed. They would ask for more 6NF and it was easy to provide (5NF to 6NF is a small step; 0NF to anything, let alone 5NF, is a big step; an organised architecture is easy to extend).

One major difference between OLTP and OLAP, the basic justification of separate OLAP platform software, is that the OLTP is row-oriented, it needs transactionally secure rows, and fast; and the OLAP doesn't care about the transactional issues, it needs columns, and fast. That is the reason all the high end BI or OLAP platforms are column-oriented, and that is why the OLAP models (Star Schema, Dimension-Fact) are column-oriented.

But with the 6NF tables:

• there are no rows, only columns; we serve up rows and columns at same blinding speed

• the tables (ie. the 5NF view of the 6NF structures) are already organised into Dimension-Facts. In fact they are organised into more Dimensions than any OLAP model would ever identify, because they are all Dimensions.

• Pivoting entire tables with aggregation on the fly (as opposed to the PIVOT of a small number of derived columns) is (a) effortless, simple code and (b) very fast

  

What we have been supplying for many years, by definition, is Relational Databases with at least 5NF for OLTP use, and 6NF for OLAP requirements.

• Notice that it is the very same science that we have used from the outset; to move from Typical un-normalised "databases" to 5NF Corporate Database. We are simply applying more of the proven science, and obtaining higher orders of functionality and performance.

• Notice the similarity between 5NF Corporate Database and 6NF Corporate Database

• The entire cost of separate OLAP hardware, platform software, ETL, administration, maintenance, are all eliminated.

• There is only one version of the data, no update anomalies or maintenance thereof; the same data served up for OLTP as rows, and for OLAP as columns

The only thing we have not done, is to start off on a new project, and declare pure 6NF from the start. That is what I have lined up next.

What is Sixth Normal Form ?

Assuming you have a handle on Normalisation (I am not going to not define it here), the non-academic definitions relevant to this thread are as follows. Note that it applies at the table level, hence you can have a mix of 5NF and 6NF tables in the same database:

• Fifth Normal Form: all Functional Dependencies resolved across the database
  • in addition to 4NF/BCNF
  • every non-PK column is 1::1 with its PK
  • and to no other PK
  • No Update Anomalies
    .
• Sixth Normal Form: is the irreducible NF, the point at which the data cannot be further reduced or Normalised (there will not be a 7NF)
  • in addition to 5NF
  • the row consists of a Primary Key, and at most, one non-key column
  • eliminates The Null Problem

What Does 6NF Look Like ?

The Data Models belong to the customers, and our Intellectual Property is not available for free publication. But I do attend this web-site, and provide specific answers to questions. You do need a real world example, so I will publish the Data Model for one of our internal utilities.

This one is for the collection of server monitoring data (enterprise class database server and OS) for any no of customers, for any period. We use this to analyse performance issues remotely, and to verify any performance tuning that we do. The structure has not changed in over ten years (added to, with no change to the existing structures), it is typical of the specialised 5NF that many years later was identified as 6NF. Allows full pivoting; any chart or graph to be drawn, on any Dimension (22 Pivots are provided but that is not a limit); slice and dice; mix and match. Notice they are all Dimensions.

The monitoring data or Metrics or vectors can change (server version changes; we want to pick up something more) without affecting the model (you may recall in another post I stated EAV is the bastard son of 6NF; well this is full 6NF, the undiluted father, and therefore provides all features of EAV, without sacrificing any Standards, integrity or Relational power); you merely add rows.

▶Monitor Statistics Data Model◀. (too large for inline; some browsers cannot load inline; click the link)

It allows me to produce these ▶Charts Like This◀, six keystrokes after receiving a raw monitoring stats file from the customer. Notice the mix-and-match; OS and server on the same chart; a variety of Pivots. (Used with permission.)

Readers who are unfamiliar with the Standard for Modelling Relational Databases may find the ▶IDEF1X Notation◀ helpful.

6NF Data Warehouse

This has been recently validated by Anchor Modeling, in that they are now presenting 6NF as the "next generation" OLAP model for data warehouses. (They do not provide the OLTP and OLAP from the single version of the data, that is ours alone).

Data Warehouse (Only) Experience

My experience with Data Warehouses only (not the above 6NF OLTP-OLAP Databases), has been several major assignments, as opposed to full implementation projects. The results were, no surprise:

• consistent with the science, Normalised structures perform much faster; are easier to maintain; and require less data synching. Inmon, not Kimball.

• consistent with the magic, after I Normalise a bunch of tables, and deliver substantially improved performance via application of the laws of physics, the only people surprised are the magicians with their mantras.

Scientifically minded people do not do that; they do not believe in, or rely upon, silver bullets and magic; they use and hard work science to resolve their problems.

Valid Data Warehouse Justification

That is why I have stated in other posts, the only valid justification for a separate Data Warehouse platform, hardware, ETL, maintenance, etc, is where there are many Databases or "databases", all being merged into a central warehouse, for reporting and OLAP.

Kimball

A word on Kimball is necessary, as he is the main proponent of "de-normalised for performance" in data warehouses. As per my definitions above, he is one of those people who have evidently never Normalised in their lives; his starting point was un-normalised (camouflaged as "de-normalised") and he simply implemented that in a Dimension-Fact model.

• Of course, to obtain any performance, he had to "de-normalise" even more, and create further duplicates, and justify all that.

  • So therefore it is true, in a schizophrenic sort of way, that "de-normalising" un-normalised structures, by making more specialised copies, "improves read performance". It is not true when the whole is taking into account; it is true only inside that little asylum, not outside.

  • Likewise it is true, in that crazy way, that where all the "tables" are monsters, that "joins are expensive" and something to be avoided. They have never had the experience of joining smaller tables and sets, so they cannot believe the scientific fact that more, smaller tables are faster.

  • they have experience that creating duplicate "tables" is faster, so they cannot believe that eliminating duplicates is even faster than that.

• his Dimensions are added to the un-normalised data. Well the data is not Normalised, so no Dimensions are exposed. Whereas in a Normalised model, the Dimensions are already exposed, as an integral part of the data, no addition is required.

• that well-paved path of Kimball's leads to the cliff, where more lemmings fall to their deaths, faster. Lemmings are herd animals, as long as they are walking the path together, and dying together, they die happy. Lemmings do not look for other paths.

All just stories, parts of the one mythology that hang out together and support each other.

Your Mission

Should you choose to accept it. I am asking you to think for yourself, and to stop entertaining any thoughts that contradict science and the laws of physics. No matter how common or mystical or mythological they are. Seek evidence for anything before trusting it. Be scientific, verify new beliefs for yourself. Repeating the mantra "de-normalised for performance" won't make your database faster, it will just make you feel better about it. Like the fat kid sitting in the sidelines telling himself that he can run faster than all the kids in the race.

• on that basis, even the concept "normalise for OLTP" but do the opposite, "de-normalise for OLAP" is a contradiction. How can the laws of physics work as stated on one computer, but work in reverse on another computer ? The mind boggles. It is simply not possible, the work that same way on every computer.

Questions ?

---

Denormalization and aggregation are the two main strategies used to achieve performance in a data warehouse. It's just silly to suggest that it doesn't improve read performance! Surely I must have missunderstood something here?

Aggregation: Consider a table holding 1 billion purchases. Contrast it with a table holding one row with the sum of the purchases. Now, which is faster? Select sum(amount) from the one-billion-row table or a select amount from the one-row-table? It's a stupid example of course, but it illustrates the principle of aggregation quite clearly. Why is it faster? Because regardless of what magical model/hardware/software/religion we use, reading 100 bytes is faster than reading 100 gigabytes. Simple as that.

Denormalization: A typical product dimension in a retail data warehouse has shitloads of columns. Some columns are easy stuff like "Name" or "Color", but it also has some complicated stuff, like hierarchies. Multiple hierarchies (The product range (5 levels), the intended buyer (3 levels), raw materials (8 levels), way of production (8 levels) along with several computed numbers such as average lead time (since start of the year), weight/packaging measures etcetera etcetera. I've maintained a product dimension table with 200+ columns that was constructed from ~70 tables from 5 different source systems. It is just plain silly to debate whether a query on the normalized model (below)

select product_id
  from table1
  join table2 on(keys)
  join (select average(..)
          from one_billion_row_table 
         where lastyear = ...) on(keys)
  join ...table70
 where function_with_fuzzy_matching(table1.cola, table37.colb) > 0.7
   and exists(select ... from )
   and not exists(select ...)
   and table20.version_id = (select max(v_id from product_ver where ...)
   and average_price between 10 and 20
   and product_range = 'High-Profile'

...is faster than the equivalent query on the denormalized model:

select product_id
  from product_denormalized
 where average_price between 10 and 20
   and product_range = 'High-Profile';

Why? Partly for the same reason as the aggregated scenario. But also because the queries are just "complicated". They are so disgustingly complicated that the optimizer (and now I'm going Oracle specifics) gets confused and screws up the execution plans. Suboptimal execution plans may not be such a big deal if the query deals with small amounts of data. But as soon as we start to join in the Big Tables it is crucial that the database gets the execution plan right. Having denormalized the data in one table with a single syntetic key (heck, why don't I add more fuel to this ongoing fire), the filters become simple range/equality filters on pre-cooked columns. Having duplicated the data into new columns enables us to gather statistics on the columns which will help the optimizer in estimating the selectivities and thus providing us with a proper execution plan (well, ...).

Obviously, using denormalization and aggregation makes it harder to accomodate schema changes which is a bad thing. On the other hand they provides read performance, which is a good thing.

So, should you denormalize your database in order to achieve read-performance? Hell no! It adds so many complexities to your system that there is no end to how many ways it will screw you over before you have delivered. Is it worth it? Yes, sometimes you need to do it to meet a specific performance requirement.

Update 1

PerformanceDBA: 1 row would get updated a billion times a day

That would imply a (near) realtime requirement (which in turn would generate a completely different set of technical requirements). Many (if not most) data warehouses does not have that requirement. I picked an unrealistic aggregation example just to make it clear why aggregation works. I didn't want to have to explain rollup strategies too :)

Also, one has to contrast the needs of the typical user of a data warehouse and the typical user of the underlaying OLTP system. A user looking to understand what factors drive transport costs, couldn't care less if 50% of todays data is missing or if 10 trucks exploded and killed the drivers. Performing the analysis over 2 years worth of data would still come to the same conclusion even if he had to-the-second up-to-date information at his disposal.

Contrast this to the needs of the drivers of that truck (the ones who survived). They can't wait 5 hours at some transit point just because some stupid aggregation process has to finnish. Having two separate copies of the data solves both needs.

Another major hurdle with sharing the same set of data for operational systems and reporting systems is that the release cycles, Q&A, deployment, SLA and what have you, are very different. Again, having two separate copies makes this easier to handle.

---

By "OLAP" I understand you to mean a subject-oriented relational / SQL database used for decision support - AKA a Data Warehouse.

Normal Form (typically 5th / 6th Normal Form) is generally the best model for a Data Warehouse. The reasons for normalizing a Data Warehouse are exactly the same as any other database: it reduces redundancy and avoids potential update anomalies; it avoids built-in bias and is therefore the easiest way to support schema change and new requirements. Using Normal Form in a data warehouse also helps keep the data load process simple and consistent.

There is no "traditional" denormalization approach. Good data warehouses have always been normalized.

---

Should not a database be denormalized for reading performance?

Okay, here goes a total "Your Mileage May Vary", "It Depends", "Use The Proper Tool For Every Job", "One Size Does Not Fit All" answer, with a bit of "Don't Fix It If It Ain't Broken" thrown in:

Denormalization is one way to improve query performance in certain situations. In other situations it may actually reduce performance (because of the increased disk use). It certainly makes updates more difficult.

It should only be considered when you hit a performance problem (because you are giving the benefits of normalization and introduce complexity).

The drawbacks of denormalization are less of an issue with data that is never updated, or only updated in batch jobs, i.e. not OLTP data.

If denormalization solves a performance problem that you need solved, and that less invasive techniques (like indexes or caches or buying a bigger server) do not solve, then yes, you should do it.

---

First my opinions, then some analysis

Opinions
Denormalisation is perceived to help reading data because common use of the word denormalisation often include not only breaking normal forms, but also introducing any insertion, update and deletion dependencies into the system.

This, strictly speaking, is false, see this question/answer, Denormalisation in strict sense mean to break any of the normal forms from 1NF-6NF, other insertion, update and deletion dependencies are addressed with Principle of Orthogonal Design.

So what happens is that people take the Space vs Time tradeoff principle and remember the term redundancy (associated with denormalisation, still not equal to it) and conclude that you should have benefits. This is faulty implication, but false implications do not allow you to conclude the reverse.

Breaking normal forms may indeed speed up some data retrieval (details in analysis below), but as a rule it will also at the same time:

• favour only specific type of queries and slow down all other access paths
• increase complexity of the system (which influences not only maintenance of the database itself, but also increases the complexity of applications that consume the data)
• obfuscate and weaken semantic clarity of the database
• main point of database systems, as central data representing the problem space is to be unbiased in recording the facts, so that when requirements change you don't have to redesign the parts of the system (data and applications) that are independent in reality. to be able to do this artificial dependencies should be minimised - today's 'critical' requirement to speed up one query quite often become only marginally important.

Analysis

So, I made a claim that sometimes breaking normal forms can help retrieval. Time to give some arguments

1) Breaking 1NF

Assume you have financial records in 6NF. From such database you can surely get a report on what is a balance for each account for each month.

Assuming that a query that would have to calculate such report would need to go through n records you could make a table

account_balances(month, report)

which would hold XML structured balances for each account. This breaks 1NF (see notes later), but allows one specific query to execute with minimum I/O.

At the same time, assuming it is possible to update any month with inserts, updates or deletes of financial records, the performance of the update queries on the system might be slowed down by time proportional to some function of n for each update. (the above case illustrates a principle, in reality you would have better options and the benefit of getting minimum I/O bring such penalties that for realistic system that actually updates data often you would get bad performance on even for your targeted query depending on the type of actual workload; can explain this in more detail if you want)

Note: This is actually trivial example and there is one problem with it - the definition of 1NF. Assumption that the above model breaks 1NF is according to requirement that values of an attribute 'contain exactly one value from the applicable domain'.

This allows you to say that the domain of the attribute report is a set of all possible reports and that from all of them there is exactly one value and claim that 1NF is not broken (similar to argument that storing words does not break 1NF even though you might have letters relation somewhere in your model).

On the other hand there are much better ways to model this table, which would be more useful for wider range of queries (such as to retrieve balances for single account for all months in a year). In this case you would justify that improvement by saying that this field is not in 1NF.

Anyway it explains why people claim that breaking NFs might improve performance.

2) Breaking 3NF

Assuming tables in 3NF

CREATE TABLE `t` (
  `id` int(10) unsigned NOT NULL AUTO_INCREMENT,
  `member_id` int(10) unsigned NOT NULL,
  `status` tinyint(3) unsigned NOT NULL,
  `amount` decimal(10,2) NOT NULL,
  `opening` decimal(10,2) DEFAULT NULL,
  PRIMARY KEY (`id`),
  KEY `member_id` (`member_id`),
  CONSTRAINT `t_ibfk_1` FOREIGN KEY (`member_id`) REFERENCES `m` (`id`) ON DELETE CASCADE ON UPDATE CASCADE
) ENGINE=InnoDB

CREATE TABLE `m` (
  `id` int(10) unsigned NOT NULL AUTO_INCREMENT,
  `name` varchar(255) DEFAULT NULL,
  PRIMARY KEY (`id`)
) ENGINE=InnoDB

with sample data (1M rows in t, 100k in m)

Assume a common query that you want to improve

mysql> select sql_no_cache m.name, count(*) 
       from t join m on t.member_id = m.id 
       where t.id between 100000 and 500000 group by m.name;
+-------+----------+
| name  | count(*) |
+-------+----------+
| omega |       11 |
| test  |        8 |
| test3 |   399982 |
+-------+----------+
3 rows in set (1.08 sec)

you could find suggestions to move attribute name into table m which breaks 3NF (it has a FD: member_id -> name and member_id is not a key of t)

after

alter table t add column varchar(255);
update t inner join m on t.member_id = t.id set t.name = m.name;

running

mysql> select sql_no_cache name, count(*) 
       from t where id 
       between 100000 and 500000 
       group by name;
+-------+----------+
| name  | count(*) |
+-------+----------+
| omega |       11 |
| test  |        8 |
| test3 |   399982 |
+-------+----------+
3 rows in set (0.41 sec)

notes: The above query execution time is cut in half, but

• the table was not in 5NF/6NF to begin with
• the test was done with no_sql_cache so most cache mechanisms were avoided (and in real situations they play a role in system's performance)
• space consumption is increased by approx 9x size of the column name x 100k rows
• there should be triggers on t to keep the integrity of data, which would significantly slow down all updates to name and add additional checks that inserts in t would need to go through
• probably better results could be achieved by dropping surrogate keys and switching to natural keys, and/or indexing, or redesigning to higher NFs

Normalising is the proper way in the long run. But you don't always have an option to redesign company's ERP (which is for example already only mostly 3NF) - sometimes you must achieve certain task within given resources. Of course doing this is only short term 'solution'.

Bottom line

I think that the most pertinent answer to your question is that you will find the industry and education using the term 'denormalisation' in

• strict sense, for breaking NFs
• loosely, for introducing any insertion, update and deletion dependencies (original Codd's quote comments on normalisation saying: 'undesirable(!) insertion, update and deletion dependencies', see some details here)

So, under strict definition, the aggregation (summary tables) are not considered denormalisation and they can help a lot in terms of performance (as will any cache, which is not perceived as denormalisation).

The loose usage encompasses both breaking normal forms and the principle of orthogonal design, as said before.

Another thing that might shed some light is that there is a very important difference between the logical model and the physical model.

For example indexes store redundant data, but no one considers them denormalization, not even people who use the term loosely and there are two (connected) reasons for this

• they are not part of the logical model
• they are transparent and guaranteed not to break integrity of your model

If you fail to properly model your logical model you will end up with inconsistent database - wrong types of relationships between your entities (inability to represent problem space), conflicting facts (ability to loose information) and you should employ whatever methods you can to get a correct logical model, it is a foundation for all applications that will be built on top of it.

Normalisation, orthogonal and clear semantics of your predicates, well defined attributes, correctly identified functional dependencies all play a factor in avoiding pitfalls.

When it comes to physical implementation things get more relaxed in a sense that ok, materialised computed column that is dependent on non key might be breaking 3NF, but if there are mechanisms that guarantee consistency it is allowed in physical model in the same way as indexes are allowed, but you have to very carefully justify it because usually normalising will yield same or better improvements across the board and will have no or less negative impact and will keep the design clear (which reduces the application development and maintenance costs) resulting in savings that you can easily spend on upgrading hardware to improve the speed even more then what is achieved with breaking NFs.

---

The two most popular methodologies for building a data warehouse (DW) seem to be Bill Inmon's and Ralph Kimball's.

Inmon's methodology uses normalized approach, while Kimball's uses dimensional modelling -- de-normalized star schema.

Both are well documented down to small details and both have many successful implementations. Both present a "wide, well-paved road" to a DW destination.

I can not comment on the 6NF approach nor on Anchor Modelling because I have never seen nor participated in a DW project using that methodology. When it comes to implementations, I like to travel down well tested paths -- but, that's just me.

So, to summarize, should DW be normalized or de-normalized? Depends on the methodology you pick -- simply pick one and stick to it, at least till the end of the project.

EDIT - An Example

At the place I currently work for, we had a legacy report which has been running since ever on the production server. Not a plain report, but a collection of 30 sub-reports emailed to everybody and his ant every day.

Recently, we implemented a DW. With two report servers and bunch of reports in place, I was hoping that we can forget about the legacy thing. But not, legacy is legacy, we always had it, so we want it, need it, can't live without it, etc.

The thing is that the mess-up of a python script and SQL took eight hours (yes, e-i-g-h-t hours) to run every single day. Needless to say, the database and the application were built over years by few batches of developers -- so, not exactly your 5NF.

It was time to re-create the legacy thing from the DW. Ok, to keep it short it's done and it takes 3 minutes (t-h-r-e-e minutes) to produce it, six seconds per sub-report. And I was in the hurry to deliver, so was not even optimizing all the queries. This is factor of 8 * 60 / 3 = 160 times faster -- not to mention benefits of removing an eight hour job from a production server. I think I can still shave of a minute or so, but right now no one cares.

As a point of interest, I have used Kimball's method (dimensional modelling) for the DW and everything used in this story is open-source.

This is what all this (data-warehouse) is supposed to be about, I think. Does it even matter which methodology (normalized or de-normalized) was used?

EDIT 2

As a point of interest, Bill Inmon has a nicely written paper on his website -- A Tale of Two Architectures.

---

The problem with the word "denormalized" is that it doesn't specify what direction to go in. It's about like trying to get to San Francisco from Chicago by driving away from New York.

A star schema or a snowflake schema is certainly not normalized. And it certainly performs better than a normalized schema in certain usage patterns. But there are cases of denormalization where the designer wasn't following any discipline at all, but just composing tables by intuition. Sometimes those efforts don't pan out.

In short, don't just denormalize. Do follow a different design discipline if you are confident of its benefits, and even if it doesn't agree with normalized design. But don't use denormalization as an excuse for haphazard design.

---

The short answer is don't fix a performance problem you have not got!

As for time based tables the generally accepted pardigm is to have valid_from and valid_to dates in every row. This is still basically 3NF as it only changes the semantics from "this is the one and only verision of this entity" to "this is the one and only version of this entity at this time "

---

Simplification:

An OLTP database should be normalised (as far as makes sense).

An OLAP data warehouse should be denormalised into Fact and Dimension tables (to minimise joins).