Atomic Data

We failed. I recently attended the Knowledge Gap conference, where we had several discussions related to data modeling. We all agreed that we are in a distressful situation both concerning the art as a whole but also its place in modern architectures, at least when it comes to integrated data models. As an art, we are seeing a decline both in interest and schooling, with practitioners shying away from its complexity and the topic disappearing from curriculums. Modern data stacks are primarily based on shuffling data and new architectures, like the data mesh, propose a decentralized organization around data, making integration an even harder task.

When I say we failed, it is because data modeling in its current form will not take off. Sure we have successful implementations and modelers with both expertise and experience in Ensemble Modeling techniques, like Anchor modeling, Data Vault and Focal. There is, however, not enough of them and as long as we are not the buzz, opportunities to actually prove that this works and works well will wane. We tried, but we’re being pushed out. We can push back, and push back harder, but I doubt we can topple the buzzwall. I won’t stop pushing, but maybe it’s also time to peek at the other side of the wall.

If we begin to accept that there will only be a select few who can build and maintain models, but many more who will push data through the stack or provide data products, is there anything we can do to embrace such a scenario?

Data Whisperers

Having given this some thought I believe I have found one big issue, preventing us from managing data in motion as well as we should. Every time we move data around we also make alterations to its representation. It’s like Chinese Whispers (aka the telephone game) in which we are lucky to retain the original message when it reaches the last recipient, given that the message is whispered from each participant to the next. A piece of information is, loosely speaking, some bundle of stuff with a possible semantic interpretation. What we are doing in almost all solutions today is to, best we can, pass on and preserve the semantic interpretation, but care less about the bundle it came in. We are all data whisperers, but in this case that’s a bad thing.

Let’s turn this around. What if we could somehow pass a bundle around without having to understand the possible semantic interpretation? In order to do that, the bundle would have to have some form that would ensure it remained unaltered by the transfer, and that defers the semantic interpretation. Furthermore, whatever is moving such bundles around cannot be surprised by their form (read like throwing an exception), so this calls for a standard. A standard we do not have. There is no widely adopted standard for messaging pieces of information, and herein lies much of the problem.

The Atoms of Data

Imagine it was possible to create atoms of data. Stable, indivisible pieces of information that can remain unchanged through transfer and duplication, and that can be put into a grander context later. The very same piece could live in a source system, or in a data product layer, or in a data pipeline, or in a data warehouse, or all of the above, looking exactly the same everywhere. Imagine there was both a storage medium and a communication protocol for such pieces. Now, let me explain how this solves many of the issues we are facing.

Let’s say you are only interested in shuffling pieces around. With atomic data pieces you are safe from mangling the message on the way. Regardless of how many times you have moved a piece around, it will have retained its original form. What could have happened in your pipelines though, is that you have dressed up your pieces with additional pieces. Adding context on the way.

Let’s say your are building an integrated enterprise-wide model. Now you are taking lots of pieces and want to understand how these fit into an integrated data model. But, the model itself is also information, so it should be able to be described using some atoms of its own. The model becomes a part of your sea of atoms, floating alongside the pieces it describes. It is no longer a piece of paper printed from some particular modeling tool. It lives and evolves along with the rest of your data.

Let’s say you are building a data product in a data mesh. Your product will shuffle pieces to data consumers, or readers may be a better word, since pieces need not be destroyed at the receiving side. Some of them may be “bare” pieces, that have not yet been dressed up with a model, some may be dressed up with a product-local model and some may have inherited their model from an enterprise-wide model. Regardless of which, if two pieces from different products are identical, they represent the same piece of information, modeled or not.

Model More Later

Now, I have not been entirely truthful in my description of the data atoms. Passing messages around in a standardized way needs some sort of structure, and whatever that structure consists of must be agreed upon. The more universal such an agreement is, the better the interoperability and the smaller the risk of misinterpreting the message. What exactly this is, the things you have to agree upon, is also a model of sorts. In other words, no messaging without at least some kind of model.

We like to model. Perhaps we even like to model a little bit too much. Let us try to forget about what we know about modeling for a little while, and instead try to find the smallest number of things we have to agree upon in order to pass a message. What, similar to a regular atom, are the elementary particles that data atoms consist of? If we can find this set of requirements and it proves to be smaller than what we usually think of when it comes to modeling, then perhaps we can model a little first and model more later.

Model Little First

As it happens, minimal modeling has been my primary interest and topic of research for the last few years. Those interested in a deeper dive can read up on transitional modeling, in which atomic data pieces are explored in detail. In essence, the whole theory rests upon a single structure; the posit.

posit_thing [{(X_thing, role_1), ..., (Y_thing, role_n)}, value, time]

The posit acts as an atomic piece of data, so we will use it to illustrate the concept. It consists of some elements put together, for which it is desired to have a universal agreement, at least within the scope in which your data will be used.

  • There is one or more things, like X_thing and Y_thing, and the posit itself is a thing.
  • Each thing takes on a role, like role_1 to role_n, indicating how these things appear.
  • There is a value, which is what appears for the things taking on these roles.
  • There is a time, which is when this value is appearing.

Things, roles, values, and times are the elements of a posit, like elementary particles build up an atom. Of these, roles need modeling and less commonly, if values or times can be of complex types, they may also need modeling. If we focus on the roles, they will provide a vocabulary, and it is through these posits later gain interpretability and relatability to real events.

p01 [{(Archie, beard color)}, "red", '2001-01-01']
p02 [{(Archie, husband), (Bella, wife)}, "married", '2004-06-19']

The two posits above could be interpreted as:

  • When Archie is seen through the beard color role, the value “red” appears since ‘2001-01-01’.
  • When Archie is seen through the husband role and Bella through the wife role, the value “married” appears since ‘2004-06-19’.

Noteworthy here is that both what we traditionally separate into properties and relationships is managed by the same structure. Relationships in transitional modeling are also properties, but that take several things in order to appear.

Now, the little modeling that has to be done, agreeing upon which roles to use is surely not an insurmountable task. A vocabulary of roles is also easy to document, communicate and adhere to. Then, with the little modeling out of the way, we’re on to the grander things again.

Decoupling Classification

Most modeling techniques, at least current ones, begin with entities. Having figured out the entities, a model describing them and their connections is made, and only after this model is rigidly put into database, things are added. This is where atomic data turns things upside down. With atomic data, lots of things can be added to a database first, then at some later point in time, these can be dressed up with more context, like an entity-model. The dressing up can also be left to a much smaller number of people if desired (like integration modeling experts).

p03 [{(Archie, thing), (Person, class)}, "classified", '1989-08-20']

After a while I realize that I have a lot of things in the database that may have a beard color and get married, so I decide to classify these as Persons. Sometime later I also need to keep track of Golf Players.

p04 [{Archie, thing), (Golf Player, class)}, "classified", '2010-07-01']

No problem here. Multiple classifications can co-exist. Maybe Archie at some point also stops playing golf.

p05 [{(Archie, thing), (Golf Player, class)}, "declassified", '2022-06-08']

Again, not a problem. Classification does not have to be static. While a single long-lasting classification is desirable, I believe we have put too much emphasis on static entity-models. Loosening up classification, so that a thing can actually be seen as more than one type of entity and that classifications can expire over time will allow for models being very specific, yield much more flexibility and extend the longevity of kept data far beyond what we have seen so far. Remember that our atomic pieces are unchanged and remain, regardless of what we do with their classifications.


Two departments in your organization are developing their own data products. Let us also assume that in this example it makes sense for one department to view Archie as a Person and for the other to view Archie as a Golf Player. We will call the Person department “financial” and it additionally needs to keep track of Archie’s account number. We will call the Golf Player department “member” and it additionally needs to keep track of Archie’s golf handicap. First, the posits for the account number and golf handicap are:

p06 [{(Archie, account number)}, 555-12345-42, '2018-01-01']
p07 [{(Archie, golf handicap)}, 36, '2022-05-18']

These posits may live their entire lives in the different data products and never reside together, or they could be copied to temporarily live together for a particular analysis, or they could permanently be stored right next to each other in an integrated database. It does not matter. The original and any copies will remain identical. With those in place, it’s time to add information about the way each department view these.

p08 [{(p03, posit), (Financial Dept, ascertains)}, 100%, '2019-12-31']
p09 [{(p04, posit), (Member Dept, ascertains)}, 100%, '2020-01-01']
p10 [{(p06, posit), (Financial Dept, ascertains)}, 100%, '2019-12-31']
p11 [{(p07, posit), (Member Dept, ascertains)}, 75%, '2020-01-01']

The posits above are called assertions, and they are metadata, since they talk about other posits. Information about information. An assertion records someone’s opinion of a posit and the value that appears is the certainty of that opinion. In the case of 100%, this corresponds to absolute certainty that whatever the posit is stating is true. The Member Department is less certain about the handicap, perhaps because the source of the data is less reliable.

Using assertions, it is possible to keep track of who thinks what in the organization. It also makes it possible to have different models for different parts of the organization. In an enterprise wide integrated model, perhaps both classifications are asserted by the Enterprise Dept, or some completely different classification is used. You have the freedom to do whatever you want.


Atomic data only works well if the data atoms remain unchanged. You would not want to end up in a situation where a copy of a posit stored elsewhere than the original all of a sudden looks different from it. Data atoms, the posits, need to be immutable. But, we live in a world where everything is changing, all the time, and we are not infallible, so mistakes can be made.

While managing change and immutability may sound like incompatible requirements, it is possible to have both, thanks to the time in the posit and through assertions. Depending on if what you are facing is a new version or a correction it is handled differently. If the beard of Archie turns gray, this is a new version of his beard color. Recalling the posit about its original color and this new information gives us the following posits:

p01 [{(Archie, beard color)}, "red", '2001-01-01']
p12 [{(Archie, beard color)}, "gray", '2012-12-12']

Comparing the two posits, a version (or natural change), occurs when they have the same things and roles, but a different value at a different time. On the other hand, if someone made a mistake entering Archie’s account number, this needs to be corrected once discovered. Let’s recall the posit with the account number and the Financial Dept’s opinion, then add new posits to handle the correction.

p06 [{(Archie, account number)}, 555-12345-42, '2018-01-01']
p10 [{(p06, posit), (Financial Dept, ascertains)}, 100%, '2019-12-31']
p13 [{(p06, posit), (Financial Dept, ascertains)}, 0%, '2022-06-08']
p14 [{(Archie, account number)}, 911-12345-42, '2018-01-01']
p15 [{(p14, posit), (Financial Dept, ascertains)}, 100%, '2022-06-08']

This operation is more complicated, as it needs three new posits. First, the Financial Dept retracts its opinion about the original account number by changing its opinion to 0% certainty; complete uncertainty. For those familiar with bitemporal data, this is sometimes there referred to as a ‘logical delete’. Then a new posit is added with the correct account number, and this new posit is asserted with 100% certainty in the final posit.

Immutability takes a little bit of work, but it is necessary. Atoms cannot change their composition without becoming something else. And, as soon as something becomes something else, we are back to whispering data and inconsistencies will abound in the organization.

What’s the catch?

All of this looks absolutely great at first glance. Posits can be created anywhere in the organization provided that everyone is following the same vocabulary for the roles, after which these posits can be copied, sent around, stored, classified, dressed up with additional context, opinionated, and so on. There is, however, one catch.


In the examples above we have used Archie as an identifier for some particular thing. This identifier needs to have been created somewhere. This somewhere is what owns the process of creating other things like Archie. Unless this is centralized or strictly coordinated, posits about Archie and Archie-likes cannot be created in different places. There should be a universal agreement on what thing Archie represents and no other thing may be Archie than this thing.

More likely, Archie would be stated through some kind of UID, an organizationally unique identifier. Less readable, but more likely the actual case would be:

p01 [{(9799fcf4-a47a-41b5-2d800605e695, beard color)}, "red", '2001-01-01']

The requirement for the identifier of the posit itself, p01, is less demanding. A posit depends on each of its elements, so if just one bit of a posit changes, it is a different posit. The implication of this is that identifiers for posits need not be universally agreed upon, since they can be resolved within a body of information and recreated at will. Some work has to be done when reconciling posits from several sources though. We likely do not want to centralize the process of assigning identities to posits, since that would mean communicating every posit from every system to some central authority, more or less defeating the purpose of decentralization.


If we are to pull off something like the data mesh, there are several key features we need to have:

  • Atomic data that can be passed around, copied, and stored without alteration.
  • As few things as possible that need to be universally agreed upon.
  • Model little first model more later, dress up data differently by locality or time.
  • Immutability so that data remains consistent across the organization.
  • Versions and corrections, while still adhering to immutability.
  • Centralized management for the assignment of identifiers.

As we have seen, getting all of these requires carefully designed data structures, like the posit, and a sound theory of how to apply them. With the work I have done, I believe we have both. What is still missing are the two things I asked you to imagine earlier, a storage medium and a communication protocol. I am well on the way to produce a storage medium in the form of the bareclad database engine, and a communication protocol should not be that difficult, given that we already have a syntax for expressing posits, as in the examples above.

If you, like me, think this is the way forward, please consider helping out in any way you can. The goal is to keep everything open and free, so if you get involved, expect it to be for the greater good. Get in touch!

We may have failed. But all is definitely not lost.

Information in Effect and Performance

Last week we had some performance issues in a bitemporal model, which by the looks of it was the result of a poorly selected execution plan in SQL Server. The reasoning behind this conclusion was that if parts of the query were first run separately with results stored in temp tables, and these later used, the issues were gone. This had me thinking though: Could something be done in order to get a better plan through the point-in-time views?

I first set about testing different methods of finding the row in effect in a unitemporal solution. In order to do so, a script was put together that creates a test bench along with a number of functions utilizing different methods. This is the script in case you would like to reproduce the test. Note that some tricks had to be employed for some methods in order to retain table elimination, a crucial feature, that may very well have skewed those results towards the negative.

The best performers in this test are the “OUTER APPLY” and “TOP 1 SUBSELECT”. We are already using the “TOP 1 SUBSELECT” variant, and they are almost tied for first place, so perhaps not much can be gained after all. That said, the execution pattern is very different between the two, so it’s hard to draw any conclusions without proper testing for the bitemporal case.

In the bitemporal point-in-time views, the rows in effect method has to be used twice. First to find the latest asserted posits, and then from those, the ones with the latest appearance time. So, I set about testing the four possible combinations of the two best approaches on one million rows in an example model. The results are summarized below (you may need to click to enlarge the images unless you have a really good monitor and incredible eye sight).

TOP 1 SUBSELECT appearance TOP 1 SUBSELECT assertion

Time to run: 8.0 seconds. This is the current approach.

OUTER APPLY appearance OUTER APPLY assertion

Time to run: 5.1 seconds. Better than current, even if the estimated cost is worse.

TOP 1 SUBSELECT appearance OUTER APPLY assertion

Time to run: 9.5 seconds. Worse than current.

OUTER APPLY appearance TOP 1 SUBSELECT assertion

Time to run: 3.9 seconds. Better than current, and lower estimated cost.


The last of the alternatives above cuts the execution time in half for the test we ran. It also has the simplest execution plan of them all. This seems promising, given that our goal was to get the optimizer to pick a good plan in a live and complex environment. I will be rewriting the logic in the generator for bitemporal models during the week to utilize this hybrid method of OUTER APPLY and TOP 1 SUBSELECT.

Temporal Complexity

Having taken a deep dive into our convenience functionalities that aim to remove most obstacles for working with temporal data, I anew “appreciated” the underlying complexities. This time around I decided to quantify these. Just how difficult is it to introduce time in a database? Is bitemporal comparatively a huge leap in complexity, as I have been touting for years without substantial proof? The answer is here.

Tracking versions is four times as difficult as not tracking anything, and adding corrections in addition makes it forty times as difficult.

To see how we got to these results, we will use the number of considerations you have to take into account as a measure. This is not exact science, but likely to be sufficiently good to produce a rule of thumb.

No temporality

When you have no intent of storing any history in your database, you will still have the following considerations. The (rough) number of things to consider are printed in parentheses before the description of the consideration.

  • (2) Your key will either match no rows or one row in the database, no prep needed.
  • (2) The value for the key will either be the same or different from the one stored.

Total: 2 × 2 = 4 considerations.

Not so bad, most people can understand some if-else logic for four cases.

Tracking versions (uni-temporal)

Stepping up and adding one timeline in order to track versions, the changes of values, many additional concerns arise.

  • (3) Your key will match no rows or up to possibly many rows in the database, some prep may be needed.
  • (2) The value for the key will either be the same or different from the one stored.
  • (3) The time of change may be earlier, the same, or later than the one stored.

Total: 3 × 2 × 3 = 18 considerations.

In other words, tracking versions is more than four times as difficult as just ignoring them altogether. Ignorance is not bliss here though, mind my word.

Tracking versions and corrections (bi-temporal)

Taking the leap, to also keep track of corrections made over time, even more concerns arise.

  • (3) Your key will match no rows or up to possibly many rows in the database, some prep may be needed.
  • (3) The value for the key will either be the same, logically deleted, or different from the one stored.
  • (3) The time of change may be earlier, the same, or later than the one stored.
  • (3) The time of correction may be earlier, the same, or later than the one stored.
  • (2) Your intended operation may be an insert or a logical delete.

Total: 3 × 3 × 3 × 3 × 2 = 162 considerations.

If you managed to pull through the 18 considerations from tracking versions, imagine nine times that effort to track corrections as well. Or, if you came from not tracking anything, prepare yourself for something requiring forty times the mental exercise.

Tracking versions, and who held an opinion about those and their certainty (multi-temporal)

I just had to compare this to transitional modeling as well, for obvious reasons.

  • (3) Your key will match no rows or up to possibly many rows in the database, some prep may be needed.
  • (5) The value for the key will either be the same, logically deleted, held with some degree of certainty, either to the value itself or its opposite, or different from the one stored.
  • (3) The time of change may be earlier, the same, or later than the one stored.
  • (3) The time of assertion may be earlier, the same, or later than the one stored.
  • (3) Your intended operation may be an insert, a logical delete, or with consideration to existing data result in you contradicting yourself or not.
  • (2) Assertions may be made by one or up to any number of asserters.

Total: 3 × 5 × 3 × 3 × 3 × 2 = 810 considerations.

That’s two hundred times more complex than most databases. It sort of makes me wonder how I ended up picking this as a topic for my research. But, here I am, and hopefully I can contribute in making everything more understandable in the end. In all fairness, many of the considerations actually have trivial outcomes, but those who do not can keep your though process going for weeks.

Thankfully, in all the scenarios above, much logic can actually be hidden from the end user, thanks to “default” rules being applied by triggers, hiding the complexity.

Modified Trigger Logic

The triggers in uni-temporal have been rewritten in order to take advantage of the performance optimizations discovered in the bi-temporal generator. At the same time, the check constraints have been removed in favor of AFTER triggers, which are more lenient (but still correct) when inserting several versions at once. Early tests indicate the following improvements:

  • Insert 1 million rows into latest view of empty anchor:
    88 seconds with old trigger logic and check constraints
    44 seconds with new logic
  • Insert another 1 million rows with 50% restatements:
    64 seconds with old trigger logic and check constraints
    46 seconds with new logic
  • Insert another 1 million rows with 100% restatements:
    37 seconds with old trigger logic and check constraints
    42 seconds with new logic

As can be seen, the performance difference is almost negligible for the new logic, regardless of the number of restatements. The only test in which the old logic performs slightly better is when every inserted row is a restatement, which is an uncommon (and probably unrealistic) scenario.

The new logic can be tested in the test version of the online modeler, now at version

New Forums

We have migrated to new forum software, since nabble was going into maintenance mode, with an uncertain future. Your user is still available if you can remember and have access to the email you used when you registered. Click “forgot password” and you will be sent instructions to reset it. Right now you have a random unguessable password.

The new forum is available here:
Anchor Forum (

We also posted a new topic on filtered indexes here:
Filtered indexes for hot stuff – Anchor Forum (

Bitemporal Generator

We have made some performance improvements to the bitemporal generator (for SQL Server) in the Anchor modeler. Code the from the generator has been running in a production environment for a while now without issues, so it should be rather safe to test out. Let us know if you find any issues.

The bitemporal generator is a subset of the concurrent-reliance-temporal generator, aimed at high performance.

Online modeler, test version:

Peridata between Data and Metadata

Somewhere in between data and metadata there is another kind of information, which we will name peridata. Perhaps you have found yourself looking at some piece of information and thinking, is this data or metadata? In this article, not only will you get a precise definition of what is what, but also a term for data living on the fringe of its classification. In order to achieve these definitions, we will turn to the posit, which is the fundamental building block of transitional modeling.


A posit essentially captures a piece of information. Here are two examples:

p1 = [{(Archie, beard)}, fluffy red, 2020-01-01]
p2 = [{(Archie, husband), (Bella, wife)}, married, 2004-06-19]

The first posit, p1, captures the information that Archie had a fluffy red beard on the 1st of January 2020. The second posit, p2, captures the information that Archie and Bella are married since the 19th of June 2004. Posits can express properties, as in p1, and relationships, as in p2. In transitional modeling, relationships are properties that require more than one thing to take on a value. Such an approach may be unfamiliar, since in most other modeling techniques there are separate constructs for properties and relationships. The proper way to read those two posits, using the notion of roles, is:

When Archie filled the beard role the value ‘fluffy red‘ appeared on 2020-01-01.

When Archie filled the husband role and Bella the wife role the value ‘married‘ appeared on 2004-06-19.

A singular thing filling a singular role gives rise to what we usually call properties or attributes, whereas a combination of things filling a combination of roles give rise to relationships. Whenever roles are filled, some value appears. In the case of Bella and Archie it could just as well have been ‘divorced’, ‘planned’, or ‘not applicable’. In fact, for the vast majority of people we could fill the roles with the relationship is ‘not applicable’, but we tend to document these only in the rare cases such posits carry valuable information.

Given the terminology of things (Archie, Bellla) and roles (beard, husband, wife), the structure of a posit can be formalized as:

posit = [
  {(thing 1, role 1), ..., (thing n, role n)},
  appearing value, 
  time of appearance

The set in the first position of the posit is called an appearance set, followed by the for that set appearing value and its time of appearance. Posits are just pieces of information and there is no requirement that they must be true. After all, there is a lot of untrue information out there and much more, maybe even most, that is uncertain to some degree. We do not want to disqualify any information from being recorded based on its certainty.

Data and Metadata

We will now make the distinction between data and metadata. Given an appearance set, if all the things it contains are not posits, then posits containing that set are classified as data. Correspondingly, given an appearance set, if at least one of the things it contains is a posit, then posits containing the set are classified as metadata. The examples given so far are data, since neither Archie nor Bella is a posit. Instead, one of the most important examples of metadata in transitional modeling is:

p3 = [{(p1, posit), (Bella, ascertains)}, 1.00, 2020-01-02]

There is no way to determine its truthfulness from a posit alone, so an additional construct is needed. An assertion is a posit that assigns a certainty to another posit. In the example above, Bella ascertains the posit about Archie’s beard, with absolute certainty on the 2nd of January 2020. This is metadata, since p1 is a posit. Assertions are subjective, and so far we only have Bella’s view of p1. Certainty is expressed by a real number in the interval [-1, 1], where 1 is being absolutely certain of what the posit is stating, 0 is having no idea whatsoever, and -1 being certain of the opposite of what the posit is stating. If you want to delve deeper into the expressiveness given by this machinery, you can read the paper “Modeling Conflicting, Unreliable, and Varying Information“.

Another common type of metadata, particularly in data warehouses, has to do with from which source posits originated.

p4 = [{(p3, source)}, The Horse's Mouth, 2020-01-01]

There could be a whole range of information related to the posit itself, like who or what recorded it, when it was entered into a database, its associated security or sensitivity, effective constraints at the time, or rules to apply in certain scenarios. These are just some examples, but all of which would be classified as metadata, because they involve a posit in their appearance sets.

Since metadata is also expressed using posits, these can be parts of appearance sets as well. For example, in p4 the assertion p3 is a part of its appearance set, so p4 is also metadata, but on a different “level” than the already metadata p3. In such a case it makes sense to distinguish these as level-1 metadata and level-2 metadata, which could be extended up to any level-n metadata. I believe that going beyond level-1 metadata is unusual in existing implementations, and that there may be few use cases that need additional levels. However, when they are needed, they are probably also very important.


While the rules separating data and metadata are clear cut, the way to tell data from peridata is less straightforward. In transitional modeling it is possible to reserve roles for particular purposes. One such example is used for classification.

p5 = [{(Archie, thing), (Person, class)}, active, 1972-08-20]

This posit tells us that Archie belongs to the Person class since 1972-08-20, using the reserved class role. Thanks to classification being expressed through posits, it is possible to disagree on these using assertions. It is also possible to have multiple classifications at once and to let classifications expire or become active at different points in time.

As you can see, there is no posit in the appearance set of p5, so it is not metadata by our previous definition. Although, the model is likely something that traditionally would have been classified as metadata. In order to distinguish this type of data from regular data, we will use the concept of reserved roles. But then, what are reserved roles? Well, you can think of them as being similar to reserved keywords in a programming language. In fact, in the examples so far, the roles positascertainsthing, and class are already reserved in transitional modeling. The roles beardhusband, and wife depend on your domain and are instead something you as a modeler will have to bring into existance.

With this we can get definitions for all three categories.

  1. If at least one of the things contained in an appearance set is a posit, then all posits with this set are classified as metadata.
  2. If at least one of the roles contained in an appearance set is reserved, then all posits with this set are classified as peridata.
  3. If neither of above applies to an appearance set, then all posits with such sets are classified as data.

Peridata exists among your data, but sort of on the fringe, given that it requires these reserved roles. Note that it is possible to have peridata for your metadata as well, when both 1 and 2 apply. Transitional modeling will come with a set of reserved roles, all of which are domain independent, but there will also be an option for end users to reserve roles of their own.


Thanks to transitional modeling, we have been able to break down what is traditionally thought of as a single metadata concept into two categories, metadata and peridata. On the fringe of your data you will find peridata, short for peripheral data, which capture such things as the classifications in your domain. Metadata is restricted to those pieces of information that explicitly talk about other pieces of information. Whether this distinction is useful remains to be seen, but it is certainly interesting. In a relational database, for example, the classifications in the modeled domain exists as a schema. Schemas are therefore peridata. Perhaps you can think of other commonly used model artifacts that fall within the scope of peridata or metadata?

On a side note, there are already some indications that the use of reserved roles can improve performance in a database engine based on posits. If you are interested in following the developement of such an engine, check out bareclad.

The Infinite Decay of Loyalty

When most businesses think of customers, they think of them as someones with which they have more than a fleeting engagement. It therefore makes sense to think of engagement lengths, or in other words, for how long a customer is a customer. If your business falls within this category, you are likely to have asked yourself how long an average customer engagement is. If you also have a valid answer to this question, based on your particular circumstances, then I congratulate you. As it turns out, the question “How long is an average customer engagement length?” is in almost all cases ill formulated and impossible to answer. All hope is not lost, however, as we shall see.

First, let us address the issue with the question itself. In any business over a certain size, there will be some customers that are loyal to the bone. They will stay with the business no matter what, until the demise of themselves or the business. Let us call this group the “eternals”. For the sake of illustration, even though not entirely mathematically correct, let these represent infinite engagement lengths. Now, remind yourself of how an average is calculated, as the sum of some engagement lengths divided by the number of customers having these lengths. If but one of your customers is an “eternal” the sum will be infinite, with your number of customers remaining finite, yielding an infinite average.

In reality, “eternals” stay for a very long but indefinite time, not infinitely long. Regardless, the previous discussion establishes that an average will be skewed to the point of uselessness or impossible to determine because of these customers. Interestingly, changing the question slightly circumvents the problem. If you instead ask “What is the median customer engagement length?”, it suddenly becomes much more approachable. Recall that the median is the value in the ‘middle’ of an ordered set of numbers. Given the engagement lengths 1, 8, 4, 6, 9, we order these by size to become 1, 4, 6, 8, 9, and conclude that 6 can be found in the middle and is therefore the median value. When the set of numbers has an even count, the median is the average of the two midmost numbers. The important feature of the median is that it is resilient to edge cases. Even if an infinite engagement length is added to the set, the median can still be calculated. This holds true as long as you do not have more than 50% “eternals” in your customer base.

The median engagement length represents the half life of your customer base. For a given cohort, say the customers signing up a certain year, after the median engagement length in years have passed, half of them are expected to remain. That is quite an understandable measure, but one problem still remains. In order to calculate the median, at least half of a cohort must have left. If the median engagement length is indeed years for your business, would you want to wait that long to figure it out? Of course not. Now this is a scenario I’ve found myself in more than once. With very little data, find a way to figure out the median engagement length. Surprisingly and somewhat happenstance, when I was looking for solutions, I stumbled upon what may be a universal pattern for how loyalty evolves over time. You see, most forecasting is done using curve fitting techniques, and finding the right equation is key. If you have only two or three points, there are lots of equations that you can apply, most of which will have very poor predictive power.

Fortunately, I happened to be at a company some 10 years ago where there were five yearly cohorts, whose development I could follow for 1, 2, 3, 4, and 5 years respectively. When plotting these the first year of every cohort aligned almost perfectly. That indicated to me that there is some universality in the behavior of loyalty. The surprising part was that for four of the cohorts, the first two points aligned, for three the first three, and so on. Now, this indicates that there is indeed some equation that can describe loyalty at this particular company. When found, it would with rather good accuracy predict the engagement lengths of whole cohorts, even brand new ones it seemed.

Looking at the shape of the curve the points were aligning to, it dropped off quite heavily in the first year, followed by successively smaller drops. The happenstance was that I recognized this type of curve. In a fortunate turn of events I had a couple of years earlier been working with calculations on the radioactivity of matter, and the beginning of this curve looked very much like exponential decay.

In exponential decay there is a fixed amount time that passes before a cohort is halved. If you restart there, and view this as a new cohort, after the same amount of time it will halve again. Using Excel goal seek (poor man’s brute forcing), with the formula below for exponential decay I was able to quickly figure out the half life of the cohorts I had at hand. Since the half life coincides with the median I was then able to answer the question “What is the median customer engagement length?” with some confidence, even if we had not passed that point in time yet.

In the formula N₀ is the original cohort size, t are the points in time at which you know the actual size N(t), and h is the half life constant you need to determine. In fact, looking at it purely mathematically, it is actually possible to determine the average engagement length as well, if it were to behave exactly like exponential decay. This is, however, again under the assumption that you have no “eternals” and that your cohort will truncate to zero customers once decay has brought it down to less than the number 1. Wikipedia also notes that behavior is better understood as long as the cohort is large.

“Many decay processes that are often treated as exponential, are really only exponential so long as the sample is large and the law of large numbers holds. For small samples, a more general analysis is necessary, accounting for a Poisson process.”

Now, some will likely find it extreme to assume that loyalty is decaying exponentially. But, if we dive a bit deeper, it actually turns out to be the most natural assumption. Let us change the approach and instead think of a customer as having a fixed probability to churn during a given time frame. For example, if we are looking at monthly cohorts, let p be the probability that a customer has churned in a month. For simplicity we assume all customers have the same probability to churn, but in reality some will be more likely and others less likely. Even so, there will be an average corresponding to the actual number of customers lost around which the individual probabilities are distributed, in some fashion. After a month we would then get that (1-pN₀ customers remain, after two months (1-p)(1-pN₀, and so on.

This is a recursive formula that produces a series. Interestingly, if we find the correct probability this series can be made to match exponential decay perfectly.

From this we can conclude that if customers have a reasonably similar probability to churn in given time frames, the end result is necessarily exponential decay. If you want to play around with this series and curve you can do so in my online workbook in GeoGebra. Given a half life h, the formula to calculate p is as follows. For example, in order to get a half life of two time periods, a churn rate of approximately 29% per period is needed.

Graphs like the one displayed by the exponential decay are called asymptotic, because as time approaches infinity the curve will approach zero. It is not hard to figure out that if the curve instead approached the number of “eternals” it would be an even better fit to the actual conditions. Changing the formula to accommodate for this is simple:

The formula is very similar to the earlier one, but now with the additional constant E, representing the number of “eternals”. Of course, this is another number not known, and the additional degree of freedom makes brute forcing their values harder, but far from impossible. The Excel Solver plugin can do multivariate goal seeks, for example.

The green curve above is using the new formula, with a likely exaggerated 20% eternals. Both of these have the half life set to two time units. Given how closely these overlap before the first halving, they are likely to be inseparable when doing curve fitting early on. They do, however, diverge significantly thereafter, so determining E should become easier shortly after the first halving. Before that, estimating E must be done through other means, like actually engaging with and talking to customers, or in the worst case, through gut feelings.

Note that in the new formula the half life pertains to the time it takes to halve the number of “non-eternals”. In order to get the new adjusted value for the constant given a desired half life, it must be multiplied by the unwieldy factor below. In the graph above the value h = 1.41504 gives an actual half life of two time units.

Assuming that all cohorts will behave like this, and that there is a recurring inflow of new customers, one can investigate the effects this has on a customer base over a longer period of time. If we start by taking the example of decaying cohorts without “eternals” and look at 15 consecutive time periods of acquisition, another surprise is in store for us.

The red curve is the sum of all the individual, gray, cohort curves, so it is in effect what the total customer base will look like. In reality customers will likely not come in bursts between each time period, but somewhat more continuously. That would just reduce the jaggedness of the curve, but it would still retain its general shape. What is particularly interesting about this shape is that it is not constantly growing, even though we adding the same number of new customers every time period. The customer base grows fast in the beginning, but then the growth stalls. This is a mathematical inevitability.

With a constant inflow of new customers, an exponential decay of loyalty will eventually stall the growth of your customer base.

If you noticed the dotted line in the graph above it is the upper bound, the largest number of customers you will ever get. This number can actually be calculated using the ratio in the rightmost part of the formula below. With the example of a 29% churn rate per time period, the largest number of customers is between three and four times a cohort size.

Over time, some customers are bound to return after a hiatus, at which point a business may view them as new again. Returning customers, even if the business has forgotten them in the meantime, are just a variation of “eternals”. The graph above is, in other words, only valid when there are no “eternals”, neither constant nor alternating. Let us therefore look at a similar graph for the more true to life example of decaying cohorts with “eternals”.

When “eternals” are part of the equation, the growth no longer stalls, and instead becomes more or less linear after an initial phase of more rapid growth. Recall that we use the likely exaggerated 20% in these examples, which is why the line is rather steep. This is, however, an indication that even a small percentage of “eternals” will make a significant difference in the development of your customer base.

Sustained growth of a customer base is only possible when some are eternally loyal.

That being said, growth cannot continue forever for other reasons. There is a limited number of people living on this planet, or more likely a limited number of people in your target market, in which there is also competition for the customers. This places an upper limit to the possible market share any business can get. Even so, understanding the mathematical fundamentals of customer base growth and applying these to your situation can yield early and important insights.

Now, let us return to the dotted line in the final graph and see if we can find its equation. First, the recursive formula will have to be adjusted for the presence of “eternals”, so that it becomes as follows.

When many such series are summed up, one for each cohort, the resulting total sum becomes the sum of the individual terms up to n.

From this the equation for the linear asymptote can be determined, and that line is described by the following equation, where t is the time passed.

With all the intellectually challenging and rather complex work done, what remains is that rather simple equation, which in essence describes the long term behavior of your customer base growth. From it, you can easily see that if E = 0 we get the simpler and constant upper bond discussed earlier. We can also see that the steepness of the asymptote is independent of your churn rate, p. Halving the churn rate, for example, will not double your customer base growth. Also, the smaller your churn rate is, the less the effect will be of reducing it further.

Both increasing the number of “eternals” and reducing the churn rate suffers from diminishing returns. A small change will result in a relatively even smaller change in growth, and the more loyal your customers become, the less the effect will be.

In the graph above, the purple growth is after halving the churn rate, compared to the blue growth. The orange growth is instead doubling the number of eternals. The long term effect of doubling the number of “eternals” is a higher sustained growth rate, and had the graph been longer it would soon have overtaken the halved churn rate. Efforts aimed to produce “eternals” are therefore more important than efforts to reduce general churn.

With all that said, there is still one parameter that we have not tinkered with. Everything so far has relied on the assumption that the inflow is constant, every cohort has the same size. For a mature business, this is not an unlikely scenario though. But, what if the cohorts themselves grow or shrink? How would that effect compare to the effects of increasing loyalty? In the graph below, the green growth has a 1% increase in the cohort size between every point. Similarly, the red growth has a 1% decrease in cohort size. Somewhat astoundingly, such a small increase will equal the effects of doubling the “eternals”. More frighteningly, with a small decrease, the growth will again almost completely stall. This places the importance of sales in a new perspective.

Efforts to produce incremental increase in customer inflow vastly outweigh efforts to increase loyalty in terms of effect on growth.

But, does this really apply to your business? I cannot answer that question with certainty, but I can say that in the original business where I discovered this 10 years ago, recent cohorts still adhere to this behavior, and old ones have not diverged from what was predicted. We, at the company where I work now, have also applied this at two other businesses in completely different domains and other stages of development. It was a bit of a long shot, but it turns out that the patterns holds true also for them. Loyalty is decaying exponentially. Now, this is the reason why I am writing this, because I am suspecting that this could be an innate and universal property of loyalty.

I know that most of you won’t go back and start doing calculations, but to those of you who do, please let me know the results!

If this indeed holds true, even within a limited scope, spreading this knowledge should prove valuable for many.

Time in Databases

Is something in your database dependent on time? If you think not, think again. I can assure you there are plenty of such things. But, as plentiful as your time-dependent objects are, as plentiful are the creative ways I’ve seen them handled. Trust me, when you screw up time, the failures of your implementation will be felt, painfully. This is, however, understandable given the complexity of time and its limited treatment in commonplace database literature. This article aims to introduce a terminology together with some best practices and considerations that should be addressed before implementing time in a database. It is inspired by the article “Kinds of Time” by Christian Kaul, and likely has significant overlaps, but provides my slightly different view.

Primary and Documentary Times

In essence there are two purposes time can serve in a database. Time can be of a primary nature or of a documentary nature. Time of a primary nature is part of your primary keys, and your database engine will, if modeled accordingly, automatically ensure temporal integrity with respect to it. Time of a documentary nature are data points that are of a time type, like a date, but that are not part of your primary keys. If you need any constraints imposed over your documentary time, you will have to build and maintain them yourself.

For integrity reasons, any primary time values must be comparable in such a way that they form a total order. Time of day, such as 12:59, cannot be used as it will repeat itself daily, giving you no option to determine if two instances of 12:59 coincided or happened in some succession. Because of this requirement, primary times are often expressed through some calendar convention, such as Julian day, Unix time, or perhaps most commonly ISO 8601, which even accommodates for leap seconds. It is worth noting that any time that is affected by daylight saving is not totally ordered. In Sweden the hour between 02:00 and 03:00 on the last Sunday of October is repeated every year. Even so and unfortunately, I see many databases here use local time as primary time.

A decent choice for a primary time would therefore be coordinated universal time (UTC). Expressed in ISO 8601, such a time looks like 2021-01-25T07:23:47.534Z. While this may look satisfactory, there is an additional concern. The precision of the data type used to store this time in the database may debilitate the total ordering. Somewhat surprisingly, and often nastily discovered, the precision of a datetime in SQL Server is 3 milliseconds. The final digit in a time expressed as above can only be 0, 3 or 7 in the database. While this particular choice is unintuitive, there is always a shortest time span that can be represented through a data type, called its chronon. For primary times, a data type with a chronon shorter than anything happening in succession is necessary to preserve the total ordering.

Given that primary times are parts of primary keys in the database and altering primary keys is normally time-consuming, the choice of data types should be made with care. Always picking the data type with the smallest chronon, such as datetime2(7) in SQL Server with a 100 nanosecond chronon, may affect performance. While it can store a time like 2007-05-02T19:58:47.1234567 it will use 8 bytes, compared to 3 bytes for the date type, if daily changes are sufficient. Keeping primary keys small should be paramount for any database designer, since smaller keys lowers total storage and increase insert and join performance.

Documentary times are not required to have a total ordering or even be temporally consistent, making it possible for versions overlapping in time. With so much leniency choices can be made with much less consideration. Naturally, there are cases when you want to impose the same restrictions to documentary times, particularly if you intend for them to behave as primary times at some point.

Particular Recurring Timepoints

There are some particular recurring timepoints of interest, and for some reason beyond my understanding there is no standardised way to express these. Some common ones are:

  • The end of time.
  • The beginning of time.
  • Indefinitely.
  • At an unknown time.

The end of time is what it sounds like, the infinite extension of time into the future. An application for this would be if you want to express a fact such as ‘I will love you forever’. Similarly, the beginning of time is the longest possible extension of time into the past. It could be applied in an expression such as ‘gravity has always been present in the universe’. Indefinitely is similar to these, but in this case we expect an actual point in time will come to pass after which a time interval is no longer open-ended. An application, with the slight but important difference from ‘forever’ is ‘I will cherish rock music until the day I die’ or ‘my hair will turn gray one day’. Finally, there is the unknown time. It can be used both for past and future events, such as ‘The price was raised, but nobody remembers when that happened’ and ‘We will raise the price the next time crops fail’.

From a storage perspective, databases normally provide one special value; NULL, that is (somewhat horrifyingly) often used for all purposes above. Practically one could possibly reason that unknown time could be used in place of indefinitely, which in turn could be used in place of the beginning and end of time. Semantically, some important nuances will then be lost. For example, the nuance lost by stating ‘I will love you until an unknown time’ may yield an entirely different outcome.

Ideally, and if your database permits user-defined types, data types which includes and separates these particular timepoints should be implemented. ISO 8601 should also be extended with ways to express these notions. There is an interesting discussion on how to express these by here, for anyone who wants to dive deeper, which suggests that standards may be coming. Regardless, you should consider how you intend to manage particular timepoints like these.

Named Timelines

Even if there is just one single time, there are many timelines. A timeline can be thought of as an interval of time (finite or infinite) over which events happen in a temporally consistent sequence. If two events can mess up each others bonds in time, such as one moving the other in time, then they definitely do not belong on the same timeline. For example, if I have an appointment in my calendar between 9:00 and 10:00 today it lives on a different timeline from the action of me, at 08:00, rescheduling it to the afternoon. Timelines can also be separated by the fact that the events they track pertain to completely different things, and it would only decrease readability and understandability to keep them together.

Borrowing the terminology of transitional modeling, following are some examples of timelines commonly discussed in computer science and database literature. There is so little consensus on the naming of these so understanding what they represent is what matters.

The Appearance Timeline

The appearance timeline contain points in time when some value was observed, became valid, or will come into effect in real life. It tracks the natural progression between values or states, both for attributes and relationships. Note that appearance timepoints may lie in the future, such as an already known price cut coming into effect on Black Friday.

In literature it is known by many different names: Valid time [Snodgrass], Effective time [Johnston], Application time [ANSI SQL:2011], and Changing time [Anchor modeling]. I also recall hearing these synonyms from forgotten sources: Utterance time, State time, Business time, Versioning time, and Statement time.

The Assertion Timeline

The assertion timeline contains points in time when some statement is subjectively assessed with respect to its certainty. In the simple case this is done by some system acting as the asserter and statements evaluating to either true or false. It is commonly used to track the correction or deletion of values or states, both for attributes and relationships. Note that assertion timepoints cannot lie in the future. If someone corrects the rebate for the upcoming price cut on Black Friday, this correction necessarily happens in the present.

In literature it is also known by many different names: Transaction time [Snodgrass], Assertion time [Johnston], System versioning time [ANSI SQL:2011], and Positing time [Anchor modeling]. I have heard less synonyms here from forgotten sources, only Falsification time and Evaluation time comes to mind.

For further reading on how to make uncertain assertions, to even being sure of the opposite, there is more information on transitional modeling in this series of articles.

The Recording Timeline

The recording timeline contains points in time at which information is stored in some kind of memory, typically when the data entered the database. This is very useful from a logging and later maintenance perspective. With it you can keep track of how quickly your database is growing on a per object basis, or revert to previous states of the database, perhaps after an erroneous load. It could have been the case that I sent all the price cuts for Black Friday into the production database but associated with the wrong products due to a faulty join.

In literature there are a couple of other names: Inscription time [Johnston] and Load date [Data Vault]. A very poor synonym I’ve seen used is Transaction time, which should be reserved for the assertion timeline alone.

The Structuring Timeline

The structuring timeline contains the point in time at which the information had a certain structure and constraints. Yes, structure and constraints change over time too. This process is referred to as schema versioning in literature, but few mention keeping a named time line for tracking when structural changes happened. If someone comes asking why there were no price cuts for Black Friday last year, you can safely assure them that ‘price cut’ was not part of your information structure at the time.

The only other name I have seen is Schema Versioning Time, but it has a too technical ring to it, in my opinion.

Unnamed Timelines

Unnamed timelines are all the points in time that do not fall within any of your named timelines. There will be values in your database that are of a time type, but that are not immediately put onto named timelines, even if the attributes themselves are named. These may be assembled onto timelines for ad-hoc purposes or they may just be used as any other descriptive attribute. A typical example would be the point of time the receipt for the stuff I bought on Black Friday was printed. You are not likely to name the timeline on which birth dates occur either.

In literature there are a couple of other names: User defined time [Snodgrass] and Happening time [Anchor]. Again, I’ve seen Transaction time used for unnamed times when the timepoint represents some event in which a transaction took place. Again, an unfortunate confusion of terminology.

Time Tracking Scope

Before implementing time in your database, you need to consider which of the timelines above and possibly others you will need, since they need to be separable in your database, possibly as different columns in the same or adjoined tables. Along with that you will also need to determine your time tracking scope. For example, is it sufficient to track changes to any part of an address or do you need to track changes of the individual parts of an address?

If tracking any change is sufficient, you can use a single point in time for the entire address. Essentially, you will be viewing a changed address, regardless of which part changed, as a new address. If you track the individual parts you will need several points in time, one for the street, one for the postal code, one for the state, and so on. In this case the same address can have different postal codes over time.

The latter approach, tracking time for every single object (attribute and relationship) can be achieved through modeling in the sixth normal form, henceforth 6NF. With it change is visible without having to make comparisons with previous rows and no data is duplicated when only a part of something is changing.

Even if you do not go as far as 6NF your time tracking scope has to be decided, since the amount of timepoints you will store depend on it. Unfortunately, in many of the source systems I regularly fetch data from, there is usually just one column named “modified date” which is documentary. In other words you can only tell something has changed and when, but not exactly what or what came before it. In these situations you can, with a proper data warehouse, provide the history the sources lost.


If you have an implementation that keeps track of both appearance and assertion timepoints, this is usually referred to as a bi-temporal implementation. The reason is that events on the appearance timeline are in a sense orthogonal to events on the assertion timeline. It is possible for the same value to appear and to be asserted simultaneously, but also at different times, so a single timepoint is not sufficient to describe both events. Furthermore, what value appears may be retroactively corrected by a later assertion. When a value appears may be also modified by an assertion. Keeping both of these on the same timeline, if you think of it as storing the date and time in a single column in a table, would cause collisions and ambiguities.

When appearances and assertions are easy to tell apart, using two different timepoints to describe these may be complex but straightforward. Problems usually arise when you are faced with a different value but nobody can tell whether it is a correction of the existing value or supposed to replace it from some point in time. This may lead to corrupt data if the wrong assumptions are made. Another issue is the fact that if you want a bi-temporal implementation with both appearance and assertion timelines treated as primary, a single table with a single primary key cannot guarantee temporal integrity. This requires careful modeling, and only a few modeling techniques have this as a “built-in” feature.


Some of the most confusing aspects of time in databases come from the use of proxying, whether deliberate or unknowingly. If we assume that I have decided to keep track of appearance, assertion, recording, and structuring timelines in my database, with 6NF time tracking scope, then I am very much all set for anything thrown at me from a querying perspective. However, that is under the assumption that all of those timepoints will be available to me when I put data into my database.

Sadly, this is often not the case. This is true both for operational systems and data warehouses. Getting information like [Using the Megastore structure as of January 5th (The database recorded on Monday 10:12:42 that ‘The manager asserted with 95% certainty on Monday at 09:15 that “The price cut will be 25% starting at midnight on Black Friday”‘)], actually never happens, yet. We do get some of the information some of the time though.

If we are in control of the database, we will always know when data is entering it. This opens up an opportunity. In the case that we do not know the assertion timepoint, say we only get “The price cut will be 25% starting at midnight on Black Friday”, we can approximate it with the recording timepoint. In this example that means missing the mark by almost an hour. As unfortunate as this is, sometimes it is the only option.

Somewhat more dangerous, but also doable, is approximating appearance timepoints with recording timepoints. Let’s say we only get “The price cut will be 25%” and we approximate it with the recording timepoint we will be dropping the price several days too early. Since recording timepoints always “happen” in the present when they come into existence, take utmost care when using it as an approximation for appearance timepoints. Still, this may sometimes also be the only option available.

Here within lies the big fallacy though. When enough approximations have been done, the different timelines become hard to distinguish, and it seems like you can use these timepoints interchangeably. This is not the case. You should always strive to get hold of the times when they are available and if proxying is necessary, and only as a last resort, then structure your loading intervals accordingly, to minimise the damage done.

Comparing Data Vault and Anchor

So far we have talked about time in databases from a theoretical perspective. There are two modeling techniques I would like to take a practical look at, taking diametrically different approaches to which timelines serve what purposes. The two techniques Anchor modeling and Data Vault are related, both being forms of Ensemble modeling, but still have many differences.

Anchor modeling utilises 6NF to provide as granular time tracking scope as possible. It designates the appearance and assertion timelines as primary for both attributes and relationships (called ties) around a concept (called anchor), while the recording timeline is documentary. Ties are attribute-like since they have a primary timeline and in that they have no identity of their own, making tie-to-tie and tie-to-attribute connections impossible, and tie-to-anchors the only option. Anchor also maintains separate metadata for the information structure in which structuring time is primary. By treating appearance and assertion timelines as primary, the database engine will ensure bi-temporal integrity. However, that needs both to be present and have functionally adequate approximations when necessary. Anchor also makes the assumption that values are exhaustive, such that an existing value cannot become NULL, and must instead be explicitly marked as “Unknown”. There no NULL values in an Anchor model.

Data Vault is similar to Anchor, but is not 6NF and instead groups attributes together (called satellites) around a concept (called hub). A single point of time is used to track all changes within a satellite, regardless of which particular attribute changed. The big difference is that Data Vault uses the recording timeline as primary for satellites. Relationships (called links) have no primary, but include a recording timepoint as documentary. Links are hub-like since they lack a primary, and can therefore have their own identities. Theoretically link-to-link and link-to-satellite connections then become possible. The implication is that relationships that change over time must be managed through other connected objects. Figuring out that some change occurred requires you to look outside of the link. Links are also, opposed to ties, always many to many, so any additional constraints have to be managed by the application layer. If appearance and assertion timelines are present in satellites or elsewhere pertain to links, they are always documentary. I do not believe Data Vault has a notion of a structuring timeline in its standard.

The advantage of Anchor is that you do not have to worry about temporal integrity after the data has entered the database. Integrity is also practically a requirement if you want to use the technique outside of data warehousing. Anchor was designed to be a general modeling technique and it is applied in several operational systems. The downside is that you need trustworthy timepoints, which can require a lot of effort and digging in the sources. Values in a source that once existed and suddenly are NULL could pose a problem if they are indeed suddenly “Unknown” and your data type does not support it to be explicitly specified. This has, in my experience, very rarely happened, and almost always the NULL means ‘deleted’, as in asserting the statement as false, which is a different thing and handled without problems. Analysts find it easy to work directly with Anchor models, thanks to it being able to serve data as it appears at or as it was asserted at without any additional work than finding the correct bitemporal time slice.

The advantage of Data Vault is that you do not have to worry at all about temporal integrity at load time. For auditing purposes, it will reproduce inconsistencies in the sources perfectly, so if you need to provide auditing and validation reports it is an excellent choice. Since Data Vault focuses specifically on data warehousing, it is also less restricted in its choice of primary timelines. However, using the recording timeline, the temporal integrity of the now documentary appearance and assertion timelines will likely have to be taken care of later. I do believe that if any business users are going to be using the data, this must be done at some point. Pushing constraints on links to the application layer has advantages if you, for example, want to prevent bigamous weddings for Christians, but allow polygamy for Mormons. The downside is that keeping consistency in a link requires more work than for a tie. In the end about the same amount of work will likely have to be done both in Anchor and Data Vault, but with additional layers in the latter. Looking at Data Vault and its choice of recording time as primary it looks like an excellent choice for a persistent staging layer, with the usually recommended Dimensional model on top as the presentable part of the data warehouse.

In my opinion both are valid options. If you like many layers, using different modeling techniques, distributing a fixed total amount of work over them, then Data Vault is a good choice. If you do not want layers, and stick to a single modeling technique, doing a fixed total amount of work for that single layer, then Anchor is a good choice. Both have been proven in practice, also for Big Data, but Data Vault has many more implementations to date.

Imprecision and Uncertainty

Going forward I am doing active research on transitional modeling, in which two other aspects of time is also considered. First there is imprecision. There is no way to measure time with perfect accuracy, so all timepoints are imprecise to some degree. In an atomic clock this imprecision is minuscule, but not insignificant. Regardless, there are events whose boundaries are hard to determine. Like when I got married. When exactly did that happen?By using fuzzy data types, intervals, or margins of error, we can actually express imprecision in databases. There are open questions on how to address the total ordering if we allow imprecise points of time in our primary timelines. Is it possible to maintain temporal integrity with imprecise values, or will we have to treat everything as documentary, and later apply some heuristics with best guesses?

The other aspect of time is uncertainty, which is not the same thing as imprecision. Certainty is a subjective measure, in which a statement is assessed with a “probability to be true”, loosely speaking. Using certainty it is actually possible to assert that you are certain of the opposite of a statement. This takes away a hard problem of storing ‘opposite values’ in a database by instead storing a negative certainty. Taking my marriage, if I look at “Lars was married on the 19th of June 2004” I can assert with 100% certainty that it is true, even if the time is imprecise enough to pin it down to a whole day. Looking at “Lars was married between 15:00 and 16:00 on the 19th of June 2004” I may actually be less certain, and assert it with 50% certainty, since I don’t exactly remember if it was one hour earlier or not. There are some open questions on when you contradict yourself if values are imprecise and you make several (vague) assertions. If values are precise, there is an exact formula by which you can calculate exactly when you contradict yourself.


Hopefully I have not made time all too confusing compared to the post of Christian that inspired me. I do believe that time in databases is a complex matter, but that should be digestible for everyone, given that we can put ourselves on some common ground. All the different terminology and poor implementations out there definitely does not help.

It’s time to treat time more seriously.

Representing Large Networks by HIERARCHYID Chunks

If you recall, I wrote about “Polymorphic Graph Queries” a while ago. This exemplified the use of HIERARCHYID to represent the topology of a small computer network. As it turns out, there is a case in which the HIERARCHYID approach will explode in both numbers and size, making them an unwieldy choice, and it’s commonly seen in large networks. There is however a way to work around that issue. As far as I can tell, the graph tables in SQL Server still do not support polymorphic queries, so this workaround should be valuable.

Assume that we have a reasonably large computer network, with say a million or more devices. Representing the entire topology of the network efficiently turns out to require a combination of HIERARCHYID and traditional relational tables. HIERARCHYID performs well all the way down from locations, through enclosures, devices, and ports or antennas to the actual communication media (fiber, ethernet, wireless). Because of the large number of things connected to this layer, this is where they become unwieldy and explode in numbers. HIERARCHYID does not work well when you have intermediate layers with comparatively massive amounts of connections. Such a scenario could easily bring you into needing billions of HIEARCHYID:s. Storage skyrockets and performance goes down the drain.

Instead, by having a traditional many-to-many table represent such layers, in which different HIERARCHYID:s are related to each other, it is possible to get the best of both worlds and achieve the ability do sub second searches through the topology. Let’s call the structure (UID, HIERARCHYID) a chunk, where the UID can typically be an integer. The relational table can then be as simple as (UID, UID) indicating that two chunks are connected, only requiring as many rows as there are connections. Polymorphic queries now need to take this into account, by first finding a number of candidate chunks, then join these through the relational table to discard ones that are not connected, which yields the final result.

A similar recursive query used for testing a relational parent-child hierarchy of the same network had to be stopped after having run for several hours. The benefit of HIERARHYID is substantial, but only if you take special care of layers with high connectivity. For small uncomplicated hierarchies, like employees and managers at a company, a traditional representation with less complexity is likely sufficient. Some alternatives can be found in “Hierarchical Data in SQL” by Ben Brumm.