BI review & help

One of the basic requirements of a modern BI system or Data Warehouse is to be non-volatile, meaning that the DWH will save the history. In this post I'll explain how to implement it in the DWH while assuming that the history can be changed and (the more interesting part) how to implement it in Analysis Services.

Before starting with history saving in the DWH, I need to say that this has nothing to do with Slowly Changing Dimensions. You can work with Type 1, 2 or 3. All you need to take care of is that the fact table will be consistent with the dimension tables. Assuming that the history can change (this is the scenario in my organization), we need to extract all the fact data every time. If you work with staging level/environment/layer you can use it to calculate which records has been changed or inserted and only upsert them to the DWH itself (it's more secure). In our case, we'll work without such staging level so we need to load all the fact data in every load to the DWH.

Loading the fact in the ETL is very simple - we'll save the datetime of the execution start, round it and put it in every row in the fact table. We can call this field CreateTime. You probably ask yourself why we need to round it. The answer is that we don't want the time dimension to have every possible value of datetime. It's huge and the users don't need that kind of accuracy. That's why I round it in my projects to half-hour, meaning that 14:27 becomes 14:00 and 14:50 becomes 14:30. Let's look at an example of such fact table:

RowNum	Dimension1	Dimension2	Measure1	Measure2	CreateTime
1	X	A	4	20	03/03/2010 15:00
2	X	B	56	19	03/03/2010 15:00
3	Y	B	2	3	03/03/2010 15:00
4	Y	B	1	10	03/03/2010 15:00
5	Y	C	7	5	03/03/2010 15:00
6	X	A	4	20	03/03/2010 15:30
7	X	B	56	40	03/03/2010 15:30
8	Y	B	2	3	03/03/2010 15:30
9	Y	B	10	10	03/03/2010 15:30
10	Y	C	7	5	03/03/2010 15:30

Here we can see two executions of the ETL: The first between 15:00 and 15:00, the second between 15:31 and 15:59. Note the differences in rows 2 vs. 7 and 4 vs. 9. This means that the source table(s) has changed between the executions.

In the DWH, we create foreign keys from Dimension1 and Dimension2 to their dimension tables. We'll also create FK between CreateTime and the time dimension. I won't get into the time dimension design here, you can read about it a past post. I assume that the time dimension has the hierarchy Year -> Month -> Day -> Hour -> Half-Hour. It seems that now we have a good fact table for use in Reporting Services (or any other reporting tool). Create a parameter called Extraction Time/ETL Time and assign it values from distinct on CreateTime. The user can pick a date and the report will filter on this date. Perfect.

The problematic part starts in Analysis Services. Only one extraction time can be the current member. There is no sense in summing or aggregating more than one extraction time. There are two ways to deal with it:

The wrong way - I'm writing this down so you'll learn from this design mistake. We created a calculated member which takes the last extraction time:

Create Member CurrentCube.[Extraction Time]. [Calendar].[All].[Last Data Time]
AS
Tail ([Extraction Time].[Calendar].[HALF HOUR].Members,1).Item(0),
Visible = 1;

In every Panorama view, we sliced the Extraction time on the [Last Data Time] member, which will show the user the data of the last ETL run. Our basic assumption was that the user wants to see the current data and when he'll want to see the historical data he will pick another member from the Extraction time dimension. This is the point we missed and it soon  enough bumped into our face.

We forgot that every system, even BI system which is built for analysts should be intuitive and idiot-proof. The users picked members from the day level in the Extraction time dimension. When we asked them why they did it they said that "we wanted to see the data from 03/03/2010 and suddenly the numbers got bigger!". That happened, of course, because the member 03/03/2010 is the sum of 03/03/2010 15:00 and 03/03/2010 15:30. The solution is that no matter which member the user picks in the Extraction time dimension, he should see only one leaf (half-hour level) member at a time. This takes us to

The right way - We set the [Last Data Time] member to be visible = 0 (I kept it for debug purposes) and I written MDX script which I call NonLastEmptyLeaf. This script makes sure that no matter which member in the Extraction time hierarchy will be picked, the user will be sliced only on the last non empty leaf. For example, if the user picks 03/03/2010 he will see the data of 03/03/2010 15:30. This is what will happen also when he'll slice on 03/2010, 2010 and even the [All] member. This is the MDX script:

Freeze([Extraction time].[Calendar].[All]);
Scope(
  Descendants (
   [Extraction time].[Calendar]. [All],
   0,
   SELF_AND_AFTER
  )
);
  this = Tail (
   Filter(
    Descendants(
     [Extraction time]. [Calendar].CurrentMember,
     [Extraction time].[Calendar].[HALF HOUR],
     LEAVES
    ),
    [Measures].CurrentMember > 0
   )
  ).Item(0);
End Scope;

Now, there is no meaning of [Last Data Time] member because slicing on the [All] member will show the current data, meaning that no slicing at all will show the current data because [All] is the default member.

To conclude, history browsing can be dangerous if we let the users the option to slice on more than one data time. The open question that I have is how can I prevent the user from picking more than one member in this hierarchy in Panorama views? Should I even try to do that or should I count on his minimal common sense?

Two weeks ago I did a talk about SCD. Here is the presentation.
I think that even if you're doing BI (design or development) for quite a while you'll find this presention resourceful and maybe you'll learn a thing or two...
Have a nice learning!

Last August, I've seen an Informatica webinar featuring Dr. Ralph Kimball (you can see it here). The topic of the webinar was "Reverse the flow of Data Warehouse Design to Ensure Success". In order to make long things short, what he says is that in the DWH's first steps you don't and can't know what are the full requirements. This is quite problematic, because the first steps of the DWH - its design - are the most crucial steps, where you take the most important decisions which will be hard to change later. His solution is revolutionary and sounds a little bit hallucinatory - Start with the website. After you have the front-end finished, whether it's a dashboard website or just a simple tabular report, and you have the customer's approval, you can start the DWH design and you know exactly what you should do and what you should not.

Thinking of that, we decided that he's right but it's too much. The cost of building a website with a data source behind it which will be changed 100 times in a short period is very high. Instead, we're working in a defferent method - We are working with Mockups.

Our mockups are PowerPoint presentations which mock the dashboard website that will be the project's front-end. We're using a lot of buttons and hyperlinks between the different slides in order to create a feeling of a real website with navigation and interactivity. When we show it to the customer and tell him that this is what his dashboard will look like, at first he's very happy and then the new requirement's phase begin: Can I have this button here? This one more slicer there? What about replacing this chart with this one? etc. In PowerPoint presentation, making these changes is very easy and after a couple of days we have a good understanding of how our front-end will look like. Threfore, we fully understand how our DWH will look like. This is a great method, I recommend you'll start using it in your next project. Try and see for yourself!

There are some more points you need to know before starting using it:

• You're not have to use Power Point for this. Visio and even Paint Brush can assist you. Nevertheless, I recommend you to use a tool for this (remember that the mockup will be changing a lot). There are many free & open-source tools to do it.
• You may consider to have the mockup to be your contract with the customer. On the other hand, the final result may look a little bit different. You can draw everything in PowerPoint but in the reality things may be hard to implement.
• After finishing the design phase, don't go over and over to the mockup to keep it updated. Again - The reality is stronger than the first design. When the project will be finished, no-one will care about the mockup.

The BI team and the source system team has a great bond between them. When the source system's team wants to change the system's logic or the physical tables, they must inform the BI team as soon as possible, because the BI team must change their system too. Otherwise, the ETL will fail and the users will see old or no data, or in the worse case the users will see incorrect data.

On the other hand, when the BI system shows data which is not the same as the data in the source system, the responsibility of showing that the BI system is OK is on the BI team's shoulders. It needs to investigate the problem and see if the error is on the BI side or in the source system side. Maybe there is no error at all and the difference is caused by difference in the logic.

Sometimes, we (the BI team, of course) investigates the data problems and find that the error resides in the source system itself. Then, we need to make sure that the error is fixed in the system, whether by the IT staff or by the power users. The important point here, and this is why I decided to write this post, is that the change must be done first in the development environment of the source system. Then, we'll run the ETL process and see that everything is fixed. This is important because sometimes we think that the change will make things right and it's not. That's why everything should be changed in the development environments first in both systems.

Only after we've seen that the data is fine, we can make the change in the production environments. Last tip: Make sure that the change is also made in the source system's production environment. Sometimes the time interval between these events is long and the source system's team just forgets all about it. Write it down as a task before going to production.

Last Week we had a funny case which took us much time to understand.

In our development environment, we like to have the same source data for a long period of time. This way we're familiar with our data and we can expect the numeric results in the higher layers (OLAP, dashboard, etc.). When we have a stable environment and we think that our system gives the correct data, we ask our infrastructure team to refresh our sources, meaning that they bring new production data to the development environment.

The situation that we had last week is that we refreshed our sources and from that point the data didn't make any sense. We asked the infrastructure team if they're sure they refreshed the data and they said they are sure. We checked the system and the data many times and seen that the data is different from what we had but still it didn't make any sense.

After much frustration, one of us made a suggestion to write sql query in this form:
select max(some_date_field) from fact_table

Then we were all amazed. After much time of sisyphean work we realized that our data is not new at all. Moreover - our old data was old too. The maximum date in our main fact table was half a year ago. The reason was that we took our data from a legacy system which runs a job that exports their tables into files, and our infrastructure team imports these files into our source tables. The problem was that no-one ran the job so the files themselves were old, meaning that our refresh activity means nothing.

Conclusion: Never trust your sources. Always run a sql query like mentioned above to check that you have recent data and that the data you think you have is the data you have. You can think of more queries that validates that what you have is what you expect to have. Run them every morning to make sure you're ground is stable. It's only a matter of seconds and it's worth it.

May the source be with you.

Very easy. In Sql Server:
Select top n * from my_table
In Oracle:
Select * from my_table where rownum <= n

It can be very useful in many many cases. For example, you're designing DWH over a system and you're looking at a certain field in one of its tables. You want to know which values this field contains, but fetching "select distinct my_field from my_table" takes too much time. Instead, if you believe know that the data is well distributed, you can use "select distinct my_field from my_table where rownum <= n". Use 1000 for n in the first trial and add one zero in the end of the number n every time until you got a query that takes too much time than you want to wait. after you got the n you can live with, can use the values you have in your query result.

There are some things you can only learn in the hard way. It didn't happenned to me personally but to my team friends, so I consider it as it is my bad.

We upgraded our ETL tool - Informatica, from version 8.5 to 8.6. We had to run some tests to see that the results are the same. So, what we did is to save the result table from 8.5 in Excel, save the result table from 8.6 in Excel and then compare them using Excel-built-in functions. The only problem is we didn't pay attention to the places where zero and null interchanged. This happened because the two versions act differently where null values take places in aggregation functions. For example, when there is a sum function and it aggregates only on null values. In one version the output is zero and in the other the output is null...

This can also happen in other tools and technology. For example, in OLAP cubes the difference between zero and null is the difference between seeing the member of the dimension on the screen and not knowing of its existence.

For conclusion, always be aware to this point and don't forget to check it.

This month we're really busy with a very important project and a short schedule. This made me think of ideas for agile development for BI, but I'll leave it for other time for now. In order to make us better BI developers, I decided to take one Pragmatic Programmer principle and use it. I took one of the most important (for my opinion) principles - DRY (Don't Repeat Yourself). The DRY principle says that "Every piece of knowledge must have a single, unambiguous, authoritative representation within a system". In classic programming it's simple to use: Use methods and generic classes to implements logic that will repeat itself in the project. But how do you do it in BI development? Here are some ideas I thought and even implemented some of them in my environment. Every layer/step in the BI development has it's own bulletin. I'll be happy to hear more from you.

• First of all - use functions in your DataWarehouse's database. Do it as much as you can. Do not repeat any logic twice or more, no matter if it's in procedures, views or even CLR functions.
• We all have much logic that repeats itself in the ETL process. For example, we found ourselves doing over and over the next process: When we build a fact table, we take every cell that points to a dimension table by a foreign key and "looking up" if it's found in the dimension table. If it's not there we replace it with Undefined, UD or null. That makes us feel very bad because we feel that we're doing the same all the time and it gives us the feeling of machines rather than programmers. The solution for this problem (and many other) is to build our own tasks (in SSIS) or transformations (in SSIS & Informatica). Alberto Ferrari did a beautiful work in this field in SSIS. I'll add some transformations of my own once I'll have release-ready versions of them.
• My co-workers just loves the Calculated Member feature in the Data Source View in SSAS. In enables them to make a new column without making a view and with no touch in the underlying database. The problem here is that after a while we have a LOT of calculated members, many of them repeats themselves and when you look for logic you lost, you can look for hours in the never-ending DSV. The solution here is not using calculated members at all. Put all your logic in the database (and as I said - in functions). The only place where you should use calculated members is where you must - when you have no write permission to the DataWarehouse or when you build your DSV over an operative database and you don't have write permissions.
• The same is with Names Queries in the Data Source View in SSAS. Don't use it.
• There's much logic that you can do only in MDX. Here, the problem is that MDX scripts are defined over cubes and not over dimensions, meaning that if a dimension has MDX logic you have to repeat it in every cube's MDX script. The solution is to add the MDX programically using AMO. Every time the ETL process ends, it should run a program that takes the MDX script from a single file and place it in every relevant cube. I know it sounds a little bit wacky and I even didn't do it myself, but for what I know, it's the only solutions for DRY in MDX.

As I said, I'll love to hear your ideas about this topic.

I'll start from the bottom line: If you create your Data Warehouse and you follow the DW rules, you're life will be easy (assuming you know the semantics and the way to build a good and correct DW).
In our case, if you have a dimension and you want to make a Parent-Child hierarchy, your life will be easy if you built the dimension's table in the right way.

For example, let's look at a simple Time dimension:

Time_Key	Day	Month	Quarter	Year	Level_Num
01012008	01	01	1	2008	1
30122008	30	12	4	2008	1
01/2008		01	1	2008	2
12/2008		12	4	2008	2
Q1/2008			1	2008	3
Q4/2008			4	2008	3
2008				2008	4

As you can see, this time dimension's table contains the days 01/01/2008 and 30/12/2008 and their parents in the levels: month, quarter and year.
Now, let's say that I need to take this dimension and make it a Parent-Child table. This is very simple. Just create a view with one new column which will be the parent column. This is the new column's code (in Pseudo-SQL):
if (level_num = 1) then Month + "/" + Year
else if (level_num = 2) then "Q" + Quarter + "/" + Year
else if (level_num = 3) then Year
else null

The result:

Time_Key	Day	Month	Quarter	Year	Level_Num	Parent
01012008	01	01	1	2008	1	01
30122008	30	12	4	2008	1	30
01/2008		01	1	2008	2	1
12/2008		12	4	2008	2	4
Q1/2008			1	2008	3	2008
Q4/2008			4	2008	3	2008
2008				2008	4	null

This is it. Now, in your OLAP DB, just configure the new column as the parent and you have a Parent-Child hierarchy. In Analysis Services you even don't have to create a view. In the Data Source View you can add a named calculation and put your code there.

See? When you create your DW according to its rules, the life is easier. In this example, if you created rows for every level in the hierarchy and created a descriptive key - everything is great.
This can help you in many scenarios. For example, when you find that your dimension is not balanced then you might want to make it a parent-child, so you won't have many pseudo-levels when the only relevant member is the leaf. Otherwise, this can be very annoying to your user. In SSAS, make sure you don't use it too much, because it is bad for performance.

Although SSAS willl let you use them, some other application such as Panorama won't function properly. I'll give one example: When you perform Drillthrough in Panorama, on the fly the engine gererates a web page that will take the user to the next view. The next view will be sliced as the current view, so this web page needs to pass the dimensions parameter (meaning - the current slices). That's why this web page contains this line:

AttachParameters("Slicers","%Slicers%")

A big problem will occur if one of the sliced dimensions will be sliced on a member with a name that contains inverted commas ("). The JavaScript will result an error because there are three inverted commas in the second parameter of the line. This is only one example of what can happen in a BI consumer program if you'll use special characters in member names. So - be careful not to pass these characters from the DW (build the ETL so it will drop these characters) or giving those names in SSAS, such as the All member name.

I feel like I don't have the right to write about it after so many bits of information were moving in the web about this subject but I would like to add my point of view (or in fact - my point of code). First of all, I must mention some of those who written about this subject before me. Mosha Pasumansky wrote a long post about it in last may. This post contains some ideas of how to come along with this problem, but none of them is perfect. In fact (as always) - there is no perfect solution for this problem. Another important source of knowledge can be found here in the MSDN forums, where Chris Webb, Thomas Pagel and others discussed it. Now, I would like to add my solution. Take it or leave it - your choice.

First of all, create a column in the time dimension that will be the current day indicator. Thanks to the Data Source View or the UDM approach of SSAS 2005 you don't have to change the relational table itself. Just add a named calculation in the DSV with expression that will be 1 for the row of the current day and null or zero for the others. The expression syntax itself depends on the underlying DB so I won't write it here, but it's very simple. Add this column as an attribute (let's call it CurrentDayInd) in the dimension structure and select its AttributeVisible propertiey to false. That's because we don't need such Attribute Hierarchy in our time dimension. After that, create a new User Hierarchy (you can call it Current Day), where the first level in it will be CurrentDayInd and after it place the day (key) attribute.

Now, what we have here? We have a hierarchy (Current Day) with two members - 1 and zero. The 1 member has only one child which is the current day. Link that member to your regular hierarchy (it's called YSQMD at my place) and here you have it. For example, you can use it that way in the MDX script:

Create Set [Last 30 Days] as
LinkMember([Time].[Current Day].[CurrentDayInd].&[1].Children.Item(0),[Time].[YSQMD])
:
LinkMember([Time].[Current Day].[CurrentDayInd].&[1].Children.Item(0),[Time].[YSQMD]).Lag(30);

I think that this solution is good and elegant. First of all, you don't have to use external functions such as Now(). The second pro is that we use the native OLAP mechanism which saves time and makes the queries run faster. Believe me, you'll feel the difference with big cubes. The last thing is that this solution is easy to understand (at least I think so) and it is easy for maintenance. The big con is that you have to process this dimension (and related cubes) every day. I don't think that it's so bad because most of the organizations do make process every day.

 

This is the second bug I found when working with AS and Oracle as my database (you can read about the first one here). Some background about our Datawarehouse architecture before I begin to complain about Microsoft:

We cannot afford risking that our users will experience faults or crashes while we refresh our DWH, process our cubes, etc. What we do to solve this problem is called a switch: We have two schemas in our Oracle DB. While our users watch the first schema, we update the other one. Only after we finish all our load process, we switch our user's tools to see the data from the second schema. In order to implement this architecture we use Oracle's synonyms. Let's say that the users watch the fact table "sales". We have a synonym which is called sales_syn. While it's pointing to the first schema (schema_a.sales_fact), we're loading into the second schema (schema_b.sales_fact). After that, we switch the synonym so it will point to the second schema (schema_b). The users always look on views that rely on synonyms. The views never change, only the synonyms do.

The problem starts when we make Named Queries in our Data Source Views in AS. Apparently, AS looks inside the view that we enterd into the DSV, finds the target of the synonym and saves it. Even after we make the switch (the synonym points on the other schema), the named query will be pointing on the first one. All our efforts to edit the named query have failed. When we open it again, we will always find that the first schema is "burned" there and cannot be changed.

The solution we chose is simply avoid using named queries. If we need a simple calculation we can add a Named Calculation in the Data Source View and if we need a complex view over our fact table, we write it ourself in the DB. This solutions breaks a little bit the main point of the data source view (designated place for all the logic of the DWH), but it is the best solution we could think of right now.

Dear Microsoft developers - it seems that you tried to be smart and look into our Oracle objects in order to enhance the multidimensional database's performance. Next time, please think twice before you do.

I believe that every BI developer seen this in many Data Warehouses: Boolean Dimensions. As you may guess, boolean dimension is a dimension with only two members and of course with no hierarchy. For example: cash/credit card in sales cube, exists/not exists in inventory cube, etc. If you haven't seen this phrase before - relax - I just invented it.
Now, the question is what to do about these dimensions:
a. Include them in the ETL process or just leave it as is?
b. If you put it in the ETL - how would you implement it?

Here's what I did in my project. You may disagree with me and I would like to see other approaches too.
a. Yes, I included it for some reasons. As every Pragmatic Programmer knows, everything can be changed so do not assume anything as globally-static. This rule takes place in here: Boolean dimensions may grow and have more members. For example, in the sales cube I mentioned above, maybe there will be another way to pay such as exclusive card of the shop (There is a network here in Israel who has it). Even male/female boolean dimension may have an Unknown member. So never exclude these dimensions from your ETL process. Wait - one more thing. You may think: Why interrupt my ETL process with these silly dimensions? If they'll grow up I'll add them to the process. As an answer think about the timings: You can never know how much time the dimension's ETL will take (although it will be very small), so in order to stay away of surprises - include it in your ETL process. just for case.
b. I implemented it as two hard-coded expressions and sent them to union. The result of this union will enter directly to the target table. In Informatica, the mapplet can't start without source table so just put a dummy table with only one row and connect it to the expression items. Why only one row? If the table will contain more than two rows then the Informatica server will consider the process as failed one.

As I said, I'll be happy to read other approaches other than mine.

I really think that the time dimension is the most complex dimension in 90% of the DWHs. The complexity is in two places: In the DWH design and also in Analysis Services (or any other BI tool).
First of all - why we didn't take the already-made Server Time Dimension which exists in SSAS 2005? For two reasons: The first is that the Project Real guys do not recommend using it (you can find their SSAS article here). The second is that we wanted to have some features that are not available in the server time dimension, such as Hebrew date. In a matter of fact, even if we didn't have such feature we still would build the time dimension ourselves because it's giving you much more control over the dimension. For example, you can always add some new attributes which Microsoft developers didn't think about.
I started myself to build the time dimension in excel. I figured out that this mission is little more complex that I thought it would. Most of the functions I wrote were simple, but there were some complicated ones. So here are some tips for you if you want to build your time dimension using Excel:

• If you want to week number for every date, do not write the function yourself... Excel has function called weeknum. If you don't have it just add the function toolbox which has it (I can't recall its name right now. check in excel help).
• If you want to have records for every level in your hierarchy (not only for days), put every level in its own excel file (not excel tab). It will help you later when you will transfer it to your DB.
• Check yourself. Pick randomly some dates and check that all of its record has correct data.

After building the excel files I needed to transfer it to my Oracle server. I used SSIS because I didn't want to wait for my DBA to copy these files into the Informatica server (it can't use my the local files, it has to be in its server. SSIS can use local files). This also was a little tricky. First of all, close excel when running the SSIS packages, otherwise it will fail. Second, when moving the non-leaf levels, go into the columns section in the destination box and erase the irrelevant columns. It will reduce the chance for errors. Finally, click on the source box and click on "Show advanced editor". Enter the source's output columns options and define properly the columns' data types. This also will reduce the chance for errors.

I had a little bit of an argument with my DBA about how should the time dimension be. I think that the time dimension does not have to be processes at all. My time dimension is from 1960 until 2020, so no daily ETL is required. She says that all the logic has to be in Informatica so I need to develop a mapping for this dimension. I think that we both are right and that's because that in ideal world she is right. In every developers team, all the BL has to be in one place. But we don't have much time (the deadline is very close) so I won't spend the time building more mapping in Informatica when I have the time dimension already made in excel.

Maybe someday I will have the time to do this. Maybe not.

I guess that this will not be my last post on this subject, but I want to start sharing some thoughts and tips from my experience when designing and building DWH. In this post I will focus on the fact & dimensions tables relationship in terms of data completeness (if you wonder what it is, read on).

Before you start to design the DWH, sit and talk with the people who built the systems which you take your data from, including the DBA. For every table, ask them what is the primary key (it's NOT always defined properly in the DB), then ask them again and then ask them if they are sure. It happened to me that I discovered that the systems guys were wrong about their DB's primary keys.
The same thing is about Foreign keys and here you should be even more careful. Even if they claim so, check yourself that every foreign key in the fact table is placed correctly in the dimension table, especially when the fact table has far history records. Sometimes system developers or even worse - system DBAs delete records from the dimension tables that are not relevant. This will cause that these keys will still be in the fact's history records but will not be found in the dimension table, causing uncomplete relationship between the fact and the dimension table.

So far is about the part when you talk and "investigate" the system developers (the DWH design). What to do when you actually developing the DWH? First, develop the tables of your dimensions tables. Do not forget to add the primary keys in the dimensions tables and the primary and foreign keys in the fact table. Then develop the ETL processes and go for the dimensions first.  If you know that the dimension has completeness problems with the fact table that you will develop later (you talked with the system developers, remember?), add UNDEFINED (UD key) record for the dimension table. Later, when developing the fact table's ETL process, make Join with the dimension table and check that the records' foreign key exists there. If not - change the key to UD. In SSIS and Informatica (and I guess that also in other products I don't know, such as DataStage) you can use Lookup instead of Joiner if the dimension table is less that 1G records. That will optimize the ETL process. After you developed all your ETLs, run the dimension processes. After they finish (assuming everything went OK) run the fact table's ETL process. If it succeeded you can go and have a drink. If not - check what went wrong. If you want to know which keys didn't showed up in the dimension table and causes the incompleteness problem, you can disable (not delete) the foreign key from the fact table and run the process again. Then, with a simple SQL query, check which foreign keys don't exist in the dimension table. Go back to your ETL design and check what you did wrong. As I pointed before, in this step you might be very angry at the system developers...

That is all for now. As I said, I assume that more ideas will come on in the future.