Jos, my co-author for
the "Building Pentaho Solutions" book just pointed me to a recent article by Jeff Prenevost entitled
"The Problem with History".
Abstract
Jeff's topic, loading a hybrid Type 1 / Type 2
slowly changing dimension table is related to
data warehousing but maybe of interest outside of that context as well.
As it turns out, the particular problem described by Jeff is non-trivial, but can be solved quite elegantly in a single SQL statment. This may be a compelling alternative to the multi-step, multi-pass solution proposed in his article.
Type 1 and Type 2 Slowly Changing Dimensions
In his article, Jeff describes a method to load a slowly changing dimension (SCD) table from an
audit trail. This would be quite straight forward in case we are dealing with a
Type 2 slowly changing dimension. In that case, each row in the audit trail would also yield one row in the dimension table. If the dimension would be a
Type 1 slowly changing dimension, the matter would only be slightly more complicated - in this case only the most recent version of each object would be loaded into the dimension table.
A Hybrid SCD Type 1/2
The interesting thing about the problem described in the article is that the dimension table is a hybrid Type 1 / Type 2 dimension. That is, for some attributes, history needs to be tracked in the dimension table (Type 2 attributes), whereas only the most recent data is required for other attributes (Type 1 attributes).
Sample Data
To make things tangible, here's a sample Employee audit trail:
| SCD Type 1 | SCD Type 2 |
+--------+--------+-------------+--------+-------+------------+------------+
| empkey | name | ssn | gender | state | valid_from | valid_to |
+--------+--------+-------------+--------+-------+------------+------------+
| 14 | Jo | 323-10-1116 | F | MI | 1998-12-03 | 1998-12-27 |
| 14 | Jo | 323-10-1119 | F | MI | 1998-12-28 | 2005-04-22 |
| 14 | Joe | 323-10-1119 | F | MI | 2005-04-23 | 2005-08-07 |
| 14 | Joseph | 323-10-1119 | M | MI | 2005-08-08 | 2006-02-12 |
| 14 | Joe | 323-10-1119 | M | MI | 2006-02-13 | 2006-07-04 |
| 14 | Joseph | 323-10-1119 | M | NY | 2006-07-05 | 2006-12-24 |
| 14 | Joseph | 323-10-1119 | M | MI | 2006-12-25 | NULL |
| 15 | Jim | 224-57-2726 | M | IL | 2002-01-16 | 2004-03-15 |
| 15 | James | 224-57-2726 | M | IL | 2004-03-16 | 2007-06-22 |
| 15 | James | 224-57-2726 | M | IN | 2007-06-23 | 2007-08-31 |
+--------+--------+-------------+--------+-------+------------+------------+
The data shows the history for the employees with
empkey 14 and 15. All rows with the same
empkey value form a time line of data change events. Each row represents a change event, updating some of the employee's data. The
valid_from and
valid_to columns are used to record when the data change event occurred, so the values in these columns change for each row. For the other columns, I used
bold markup to make it easier to spot the change.
SCD Type 1 and 2 Attributes
In Jeff's article, the columns
name and
ssn end up as SCD Type 1 attributes, and the columns
gender and
state as SCD Type 2 attributes. I used color highlighting to mark up groups of consecutive rows where the values for the
gender and
state columns did not change. After loading the dimension table we must end up with one row for each such group, capturing all change events for these columns. The
ssn and
name must always contain the most recent data. I highlighted the most recent row using grey markup.
Resulting Dimension Table
The data for the resulting dimension table would look like this:
+--------+--------+--------+-------------+--------+-------+------------+------------+------------+
| dw_key | empkey | name | ssn | gender | state | valid_from | valid_to | is_current |
+--------+--------+--------+-------------+--------+-------+------------+------------+------------+
| 1 | 14 | Joseph | 323-10-1119 | F | MI | 1998-12-03 | 2005-08-07 | N |
| 2 | 14 | Joseph | 323-10-1119 | M | MI | 2005-08-08 | 2006-07-04 | N |
| 3 | 14 | Joseph | 323-10-1119 | M | NY | 2006-07-05 | 2006-12-24 | N |
| 4 | 14 | Joseph | 323-10-1119 | M | MI | 2006-12-25 | 9999-12-31 | Y |
| 5 | 15 | James | 224-57-2726 | M | IL | 2002-01-16 | 2007-06-22 | N |
| 6 | 15 | James | 224-57-2726 | M | IN | 2007-06-23 | 9999-12-31 | Y |
+--------+--------+--------+-------------+--------+-------+------------+------------+------------
As you can see, only the changes in the
gender and
state columns are recorded, ignoring any changes in the
name and
ssn columns. Instead, these columns get the values of the most recent change.
Jeff's Solution
Jeff's article describes a step-wise solution to this problem. To give you an idea and some context, here's a quote from the article describing the mindset in developing this approach:
While an ETL tool like Informatica or DataStage would be handy, any of this could be done fairly easily in straight SQL. We’ll also keep all the steps simple, easy to understand and discrete. It’s possible to create enormous heaps of nested SQL to do everything in one statement, but it's best to keep everything understandable. We create and label "tables," but whether you choose to actually create real database tables, or the tables are just record sets inside a flow, the process remains essentially unchanged.
The actual steps that make up the described solution can be summarized as follows:
- Rank all rows according to the timeline
- Join consecutive rows by rank in case the there is a change in the SCD attributes
- Break the joined rows back into two rows (Jeff calls this 'semi-pivot')
- Add the last and first version for each object to the set
- Add another rank number to the result set
- At this point, the rows where the new rank number is odd contain the
valid_from date, and rows with an even number contain the valid_to date. Join each odd row to its consecutive even rows.
- Join to the last version of each object to fill in for the SCD Type 1 attributes, and set a flag for the latest version of the object.
I admit I have a hard time understanding Jeff's method, and I am completely puzzled as to how he invented it. My apologies if my summary isn't exactly crystal clear - I encourage you to read
the original article, "The Problem with History" yourself if you are interested in the details.
My Alternative
First of all, let me say that I am impressed with Jeff's creativity. I would never think of this solution and I am not really sure I even understand it. I also am not categorically opposed to Jeff's initial mindset: I too think that it is sometimes better to avoid complex SQL statements in favor of a series of relatively simple ones.
That said, when I first read Jeff's article, my first impression was: "Wow...that are a lot of moving parts". My second impression was: "Darn, that's a lot of multiple passes over the same data set". All the time, I had this hunch that it wouldn't be to hard to do it in two statements (one
SELECT...INTO, one
UPDATE), or perhaps even in one (
SELECT...INTO).
Preparation
I started by setting up the test data in MySQL 5.1. I'm including the script here for the readers' convenience:
CREATE TABLE hrsource (
empkey int NOT NULL
, name varchar(10) NOT NULL
, ssn char(11) NOT NULL
, gender char(1) NOT NULL
, state char(2) NOT NULL
, valid_from date NOT NULL
, valid_to date
, PRIMARY KEY(empkey, valid_from)
);
INSERT INTO hrsource
(empkey,name,ssn,gender,state,valid_from,valid_to) VALUES
(14, 'Jo' , '323-10-1116', 'F', 'MI', '1998-12-03', '1998-12-27')
,(14, 'Jo' , '323-10-1119', 'F', 'MI', '1998-12-28', '2005-04-22')
,(14, 'Joe' , '323-10-1119', 'F', 'MI', '2005-04-23', '2005-08-07')
,(14, 'Joseph', '323-10-1119', 'M', 'MI', '2005-08-08', '2006-02-12')
,(14, 'Joe' , '323-10-1119', 'M', 'MI', '2006-02-13', '2006-07-04')
,(14, 'Joseph', '323-10-1119', 'M', 'NY', '2006-07-05', '2006-12-24')
,(14, 'Joseph', '323-10-1119', 'M', 'MI', '2006-12-25', NULL)
,(15, 'Jim' , '224-57-2726', 'M', 'IL', '2002-01-16', '2004-03-15')
,(15, 'James' , '224-57-2726', 'M', 'IL', '2004-03-16', '2007-06-22')
,(15, 'James' , '224-57-2726', 'M', 'IN', '2007-06-23', '2007-08-31');
Of course, this is not a lot of data, but it is the data from the original article, dutifully hacked into my computer's keyboard by your's truly.
Solution
Without further ado, this is my solution for selecting the dimension's table dataset:
SELECT prv.empkey
, curr.name
, curr.ssn
, prv.gender
, prv.state
, MIN(prv.valid_from) valid_from
, COALESCE(
nxt.valid_from - INTERVAL 1 DAY
, '9999-12-31'
) valid_to
, CASE
WHEN nxt.valid_from IS NULL THEN 'Y'
ELSE 'N'
END is_current
FROM hrsource curr
INNER JOIN (
SELECT empkey
, MAX(valid_from) valid_from
FROM hrsource
GROUP BY empkey
) curr1
ON curr.empkey = curr1.empkey
AND curr.valid_from = curr1.valid_from
INNER JOIN hrsource prv
ON curr.empkey = prv.empkey
LEFT JOIN hrsource nxt
ON prv.empkey = nxt.empkey
AND prv.valid_to < nxt.valid_from
AND (prv.gender,prv.state)!= (nxt.gender,nxt.state)
LEFT JOIN hrsource inb
ON prv.empkey = inb.empkey
AND prv.valid_to < inb.valid_from
AND nxt.valid_from > inb.valid_to
AND (prv.gender,prv.state)!= (inb.gender,inb.state)
WHERE inb.empkey IS NULL
GROUP BY prv.empkey
, nxt.valid_from
Now I would certainly not qualify this as a simple statement. At the same time, I feel this does not resemble the "enormous heaps of nested SQL" so dreaded by Jeff. Let me explain how it works, and you can judge for yourself.
Explanation
I will know do a step-by-step explanation of my statement. I hope it clears up any doubt you might have as to how my solution solves the problem.
SCD Type 1 attributes: Isolating the Most Recent Change
In the introduction, I mentioned that the loading the dimension table would be easy in case it was either a Type 1 or a Type 2 slowly changing dimension. Well, we know that, come what may, we will always need the most recent row for each
empkey to supply values for the Type 1 attributes
name and
ssn. So, we start by attacking that part of the problem.
The following fragment of my solution does exactly that:
FROM hrsource curr
INNER JOIN (
SELECT empkey
, MAX(valid_from) valid_from
FROM hrsource
GROUP BY empkey
) curr1
ON curr.empkey = curr1.empkey
AND curr.valid_from = curr1.valid_from
The subquery
curr1 finds us the highest value for the
valid_from column for each distinct
empkey. The
GROUP BY empkey clause ensures we get exactly one row for each distinct value of
empkey. From all rows with the same
empkey value, the
MAX(valid_from) bit finds the largest
valid_from value.
The combination
empkey, valid_from is the primary key of the employee audit trail. This means we can now use the
empkey, MAX(valid_from) pair from the
curr1 subquery to point out the most recent row for each distinct
empkey. The
INNER JOIN with the
hrsource table (dubbed
curr) does exactly that.
Grouping consecutive rows with identical Type 2 attributes
We must now deal with the second part of the problem, the SCD Type 2 attributes.
If you look back at the color highlighting in the sample data set, you may realize this is basically a grouping problem: for each distinct
empkey we need to group rows with identical values in the Type 2 columns
gender and
state. From the point of view of the dimension table, no change occurred within this group, so we should store it as one row in the dimension table, and reconstruct the change dates by taking the minimum
valid_from and maximum
valid_to values from the group.
We must be careful though: we can't simply make groups of all distinct combinations of the SCD Type 2 columns. Look for example at the rows having
14 for the
empkey column. The rows with the
valid_from values
2005-08-08,
2006-02-13 and
2006-12-25 all have the identical combination (
M,
MI) in the Type 2 columns
gender and
state respectively. However, only the first two of these belong together in a group. The row with
2006-12-25 in the
valid_from column does not belong to the group because another change event occurred on
2006-07-05, which lies between
2006-02-13 and
2006-12-25.
So, we must sharpen up the definition of the problem. It is not enough to make groups of identical combinations of Type 2 attributes: we must also demand that the rows in the group are
consecutive.
The first step in attacking this problem is generate combinations of the audit trail rows so that we have the first row of each group along the first row of the next group. The following fragment solves part of that problem by combining each row (alias:
prv) with all more recent rows (alias:
nxt) that have the same
empkey value but different values in the Type 2 columns
gender and
state:
INNER JOIN hrsource prv
ON curr.empkey = prv.empkey
LEFT JOIN hrsource nxt
ON prv.empkey = nxt.empkey -- timeline of same employee
AND prv.valid_to < nxt.valid_from -- nxt must be more recent than prv
AND (prv.gender,prv.state)!= (nxt.gender,nxt.state) -- type 2 attributes differ
(Please ignore the part I struck out - it's not relevant at this point)
Note the usage of the
LEFT JOIN. It ensures that if
prv is the most recent row, and there is by definition no row in
nxt that is more recent, the
prv row is still retained. Had we used an
INNER JOIN, we would have lost that row, messing up the result.
If we run this fragment in isolation, we can certainly see rows that pair the first row of a group with the first row of the next group. For example, if you run this query:
SELECT prv.empkey
, prv.valid_from prv_from, prv.valid_to prv_to
, nxt.valid_from nxt_from, nxt.valid_to nxt_to
, prv.gender prv_gender
, nxt.gender nxt_gender
FROM hrsource prv
LEFT JOIN hrsource nxt
ON prv.empkey = nxt.empkey
AND prv.valid_to < nxt.valid_from
AND (prv.gender, prv.state) != (nxt.gender, nxt.state)
We get results like this:
+--------+------------+------------+------------+------------+------------+------------+
| empkey | prv_from | prv_to | nxt_from | nxt_to | prv_gender | nxt_gender |
+--------+------------+------------+------------+------------+------------+------------+
| 14 | 1998-12-03 | 1998-12-27 | 2005-08-08 | 2006-02-12 | F | M |
| 14 | 1998-12-03 | 1998-12-27 | 2006-02-13 | 2006-07-04 | F | M |
| 14 | 1998-12-03 | 1998-12-27 | 2006-07-05 | 2006-12-24 | F | M |
. . . . . . . .
. . ...more rows... . . . .
. . . . . . . .
+--------+------------+------------+------------+------------+------------+------------+
The first row highlighted in green is a desired combination because the
valid_from date of
nxt is closest to that of
prv. It is this row from
nxt that marks the end of the group of rows starting with
prv. For the rows highlighted in red, the
nxt part indicates a change that occurred beyond that point and are thus not desired.
To get rid of these undesired rows, we simply have to demand that no row indicating a change in the SCD Type 2 attributes occurs between
prv and
nxt. In my query, the following fragment is responsible for that part:
LEFT JOIN hrsource inb
ON prv.empkey = inb.empkey -- time line of same employee
AND prv.valid_to < inb.valid_from -- more recent than prv
AND (prv.gender,prv.state)!= (inb.gender,inb.state) -- Type 2 columns changed
AND nxt.valid_from > inb.valid_to -- less recent than nxt
WHERE inb.empkey IS NULL -- does not exist
This fragment essentially states that there must not be any row in between
prv and
nxt that indicate a change in the SCD type 2 attributes as compared to
prv. Note that the first part of the join condition is identical to the one we used to join
nxt to
prv. The extra's are that we also ask that
inb lies before
nxt. Because we used a
LEFT JOIN, the result row is retained in case no such
inb row exists. The
WHERE inb.empkey IS NULL condition explicitly filters for these cases. By definition, this means that the
nxt part in the result row marks the end of whatever group the
prv belongs to.
Rolling Up the Rows in their Groups
Now, if we put these parts together we get something like this:
SELECT prv.empkey
, prv.valid_from prv_from, prv.valid_to prv_to
, nxt.valid_from nxt_from, nxt.valid_to nxt_to
, prv.gender prv_gender
, nxt.gender nxt_gender
FROM hrsource prv
LEFT JOIN hrsource nxt
ON prv.empkey = nxt.empkey
AND prv.valid_to < nxt.valid_from
AND (prv.gender, prv.state) != (nxt.gender, nxt.state)
LEFT JOIN hrsource inb
ON prv.empkey = inb.empkey
AND prv.valid_to < inb.valid_from
AND (prv.gender,prv.state) != (inb.gender,inb.state)
AND nxt.valid_from > inb.valid_to
WHERE inb.empkey IS NULL
Some of the results are here:
+--------+------------+------------+------------+------------+------------+------------+
| empkey | prv_from | prv_to | nxt_from | nxt_to | prv_gender | nxt_gender |
+--------+------------+------------+------------+------------+------------+------------+
| 14 | 1998-12-03 | 1998-12-27 | 2005-08-08 | 2006-02-12 | F | M |
| 14 | 1998-12-28 | 2005-04-22 | 2005-08-08 | 2006-02-12 | F | M |
| 14 | 2005-04-23 | 2005-08-07 | 2005-08-08 | 2006-02-12 | F | M |
. . . . . . . .
. . ...more rows... . . . .
. . . . . . . .
+--------+------------+------------+------------+------------+------------+------------+
Now, we still see the individual rows. However, because we have paired them with the row that definitely marks the end of the group it belongs to, we can now lump them together using
GROUP BY. This explains the final bits of the query:
GROUP BY prv.empkey
, nxt.valid_from
For each value in
empkey, this
GROUP BY clause essentially rolls up the group of rows that showed no change in the SCD Type 2 attributes until the occurrence of the
nxt row. In the
SELECT list, we can now use the
MIN function to find the date of the first row in the group.
Note that the
GROUP BY list is the thinnest one possible. The
SELECT list contains many more columns that do not appear in the
GROUP BY list. However, it is perfectly safe to do so in this case. For more details on this matter, please
read my Debunking Group By Myths article.
Glueing the SCD Type 1 and Type 2 Solutions together
We developed the solution as two separate queries - one to take care of the SCD Type 1 attributes, one for the SCD Type 2 attributes. These two are simply joined together over the
empkey. The code to do this was displayed struck out in the first query fragment that deals with SCD type 2 attributes.
Calculating the Result columns
The only part of the query that was not discussed yet is the
SELECT list:
SELECT prv.empkey
, curr.name
, curr.ssn
, prv.gender
, prv.state
, MIN(prv.valid_from) valid_from
, COALESCE(
nxt.valid_from - INTERVAL 1 DAY
, '9999-12-31'
) valid_to
, CASE
WHEN nxt.valid_from IS NULL THEN 'Y'
ELSE 'N'
END is_current
Note that we select the
name and
ssn columns from the
curr table. This ensures these columns will always reflect the data of the most recent change which is typical for SCD Type 1 attributes. The
gender and
state columns on the other hand are drawn from
prv. They reflect the change history for each individual employee.
The
valid_from column is aggregated using the
MIN function. This essentially yields the date of the first row of each group of consecutive rows having the same
empkey and values in the SCD Type 2 columns.
The
valid_to date is calculated from the
valid_from date of the first row of the next group of rows that indicates a change of values in the SCD Type 2 columns. This is the relevant code:
nxt.valid_from - INTERVAL 1 DAY
Now, to be completely fair - this is in fact cheating. It implies I trust the data to be such that the value for the
valid_to date is always exactly one day less than the value for the
valid_from date for the next row in the history. This was certainly the case with the sample data, but we cannot really rely on that. That said, it is not that hard to get the 'real'
valid_to date, but it would make this example a lot harder to understand.
Another thing about the calculation of
valid_to is the treatment of
NULL values. For each history associated with one distinct
empkey value, the last row will have
NULL here, and this is substituted by a large date instead. Dimension tables tend to avoid nullable columns and in this case the maximum value '9999-12-31' expresses that this row is the current row. Using that same logic, and exrta flag column is added to indicate whether the row is the current row.
As you can see in the screenshot, "Key" is the key field: the value of this field is scanned to determine in which output fields to put the corresponding value. There are no fields that make up a grouping: rather, we want all key/value pairs to end up in one big row. Or put another way, there is just one group comprising all key/value pairs. Finally, the grid below specifies for each distinct value of the "Key" field to which output field name it should be mapped, and in all cases, we want the value of the "Value" field to be stored in those fields. 
