I'm implementing a Data Mart following the Kimball methodology and I have a challenge with applying deltas from multiple source tables against a single target dimension.
Here's an example of the incoming source data:
STG_APPLICATION
APP_ID, APP_NAME, APP_START_DATE, CDC_HASH, ...
1, FOOBAR, 20/10/2018, MD5_XXX
STG_APPLICATION_STATUS
APP_ID, STATUS_CODE, STATUS_DESC, CDC_HASH, ...
1, SUBMITTED, "APP WAS SUBMITTED", MD5_YYY
Each of these tables (there are several others) represent a normalised version of the source data i.e. a single application can have one or more statuses associated with it.
Now then, because we only get a full alpha for these tables we have to do a snapshot merge, i.e. apply a full outer join on the current day set of records against the previous day set of records for each individual table. This is computed by comparing the CDC_HASH (a concat of all source columns). The result of this comparison is stored in a delta table as follows:
STG_APPLICATION_DELTA
APP_ID, APP_NAME, APP_START_DATE, CDC_HASH, CDC_STATUS ...
STG_APPLICATION_STATUS
APP_ID, STATUS_CODE, STATUS_DESC, CDC_HASH, CDC_STATUS...
1, AWARDED, "APP WAS AWARDED", MD5_YYY, NEW
So in this example, the first table, STG_APPLICATION did not generate a delta record as the attributes pertaining to that table did not change between daily loads. However, the associated table, STG_APPLICATION_STATUS, did calculate a delta, i.e. one or more fields have changed since the last load. This is highlighted by the CDC_STATUS which identifies it as a new record to insert.
The problem now of course is how to correctly handle this situation when loading the target dimension? For example:
DIM_APPLICATION
ID, APPLICATION_ID, APP_NAME, APP_START_DATE, APP_STATUS_CODE, FROM_DATE, TO_DATE
1, 1, FOOBAR, 20/10/2018, SUBMITTED, 20/10/2018, 12/04/2019
2, 1, NULL, NULL, NULL, AWARDED, 13/04/2019, 99/99/9999
This shows the first record - based on these two staging tables being joined - and the second record which is meant to reflect an updated version of the record. However, as previously illustrated, my Delta tables are only partially populated, and therefore I am unable to correctly update the dimension as shown here.
Logically, I understand that I need to be able to include all fields that are used by the dimension as part of my delta calculation, so that I have a copy of a full record when updating the dimension, but I'm not sure of the best way to implement this in my staging area. As shown already, I currently only have independent staging tables, each of which calculate their delta separately.
Please can somebody advise on the best way to handle this? I'm scrutinized Kimball's books on this but to no avail. And I've equally found no suitable answer on any online forums. This is a common problem so I'm sure there exists a suitable architectural pattern to resolve this.
You will need to either compare on joined records or lookup the current dimension values.
If the amount of (unchanged) data is not excessive, you could join the full snapshots of STG_APPLICATION and STG_APPLICATION_STATUS together on APP_ID until they resemble the dimension record column-wise and store those in a separate table with their CDC hash to use as previous day. You then take the deltas at this level and send the (complete) changed records as updates to the dimension.
If the amount of records in the daily update makes it impractical to join the full tables, you can take the deltas and full outer join them as you do now. Then you look up the current dimension record for this APP_ID and fill in all empty fields in the delta record. The completed record is then sent as an update to the dimension.
This solution requires less storage but seems more fragile, especially if multiple changes are possible within a day. If there are many changes, performance may also suffer. For a handful of changes in millions of records, it should be more efficient.
Related
I don't know if I even worded the question correctly, but I'm trying to create a measure that depends on what is showing in the pivot table (using PowerPivot). In the image I posted, "DealMonth" is an expression in the PowerQuery table itself that simply takes the start date of the employee and subtracts it from the month a deal was closed in. That will show how long it took for that salesperson to close the deal. "TenureMonths" is also an expression in the PowerQuery table that calculates the tenure of the person. The values populating this screenshot are coming from a total headcount measure created. What I'm trying to do is create a separate measure that will show when the "TenureMonths" is less than the "DealMonth." So if the TenureMonths is 5, then after DealMonth of 5, the value would be 0. Is this possible?
Screenshot
I should add the following information.
"DealMonth" - Comes from the FactData table
"TenureMonths" - Comes from the DimSalesStart table
These two tables are joined by name. I feel like I'm so close because I can see what I want. The second image below is a copy/paste of the pivot table result but with my edits to show what I'd want to have shown. Basically, if(TenureMonths >= DealMonth,1,0). The trouble seems to be that since they're in two different tables, I can't make it work. The rows in the fact table are transactions, but the rows in the dim table are just the people with their start and end dates.
Desired Result
This is possible with some IF([measure1]<[measure2],blank(),[measure1]), however without seeing more of the data it will be hard to guide you specifically.
However you need to create two separate measures, one for TenureMonths and one for DealMonth, depending on the data this can be done with an aggregator forumla such as sum, min, max, etc (depends if there will be more than one value).
Then reference those two measures in the formula pattern I mentioned above, and that should give you want you want.
I figured out a solution. I added a dimension table for DealMonth itself and joined to my fact table. That allowed me to do the formulas that I needed.
In Cassandra, a row can be very long and store units of time relevant data. For example, one row could look like the following:
RowKey: "weather"
name=2013-01-02:temperature, value=90,
name=2013-01-02:humidity, value=23,
name=2013-01-02:rain, value=false",
name=2013-01-03:temperature, value=91,
name=2013-01-03:humidity, value=24,
name=2013-01-03:rain, value=false",
name=2013-01-04:temperature, value=90,
name=2013-01-04:humidity, value=23,
name=2013-01-04:rain, value=false".
9 columns of 3 days' weather info.
time is a primary key in this row. So the order of this row would be time based.
My question is, is there any way for me to do a query like: what is the last/first day's humidity value in this row? I know I could use a Order By statement in CQL but since this row is already sorted by time, there should be some way to just get the first/last one directly, instead of doing another sort. Or is cassandra optimizing it already with Order By statement under the hood?
Another way I could think of is, store another column in this row called "last_time_stamp" that always updates itself as new data is inserted in. But that would require one more update every time I insert new weather data.
Thanks for any suggestion!:)
Without seeing more of your actual table, I suggest using a timestamp (or timeuuid if there is a possibility for collisions) as the second component in a compound primary key. Using this, you can get the last "row" by selecting ORDER BY t DESC LIMIT 1.
You could also change the clustering order in your schema to order it naturally for "last N" queries.
Please see examples and linked resource in this answer.
Suppose I have a large (300-500k) collection of text documents stored in the relational database. Each document can belong to one or more (up to six) categories. I need users to be able to randomly select documents in a specific category so that a single entity is never repeated, much like how StumbleUpon works.
I don't really see a way I could implement this using slow NOT IN queries with large amount of users and documents, so I figured I might need to implement some custom data structure for this purpose. Perhaps there is already a paper describing some algorithm that might be adapted to my needs?
Currently I'm considering the following approach:
Read all the entries from the database
Create a linked list based index for each category from the IDs of documents belonging to the this category. Shuffle it
Create a Bloom Filter containing all of the entries viewed by a particular user
Traverse the index using the iterator, randomly select items using Bloom Filter to pick not viewed items.
If you track via a table what entries that the user has seen... try this. And I'm going to use mysql because that's the quickest example I can think of but the gist should be clear.
On a link being 'used'...
insert into viewed (userid, url_id) values ("jj", 123)
On looking for a link...
select p.url_id
from pages p left join viewed v on v.url_id = p.url_id
where v.url_id is null
order by rand()
limit 1
This causes the database to go ahead and do a 1 for 1 join, and your limiting your query to return only one entry that the user has not seen yet.
Just a suggestion.
Edit: It is possible to make this one operation but there's no guarantee that the url will be passed successfully to the user.
It depend on how users get it's random entries.
Option 1:
A user is paging some entities and stop after couple of them. for example the user see the current random entity and then moving to the next one, read it and continue it couple of times and that's it.
in the next time this user (or another) get an entity from this category the entities that already viewed is clear and you can return an already viewed entity.
in that option I would recommend save a (hash) set of already viewed entities id and every time user ask for a random entity- randomally choose it from the DB and check if not already in the set.
because the set is so small and your data is so big, the chance that you get an already viewed id is so small, that it will take O(1) most of the time.
Option 2:
A user is paging in the entities and the viewed entities are saving between all users and every time user visit your page.
in that case you probably use all the entities in each category and saving all the viewed entites + check whether a entity is viewed will take some time.
In that option I would get all the ids for this topic- shuffle them and store it in a linked list. when you want to get a random not viewed entity- just get the head of the list and delete it (O(1)).
I assume that for any given <user, category> pair, the number of documents viewed is pretty small relative to the total number of documents available in that category.
So can you just store indexed triples <user, category, document> indicating which documents have been viewed, and then just take an optimistic approach with respect to randomly selected documents? In the vast majority of cases, the randomly selected document will be unread by the user. And you can check quickly because the triples are indexed.
I would opt for a pseudorandom approach:
1.) Determine number of elements in category to be viewed (SELECT COUNT(*) WHERE ...)
2.) Pick a random number in range 1 ... count.
3.) Select a single document (SELECT * FROM ... WHERE [same as when counting] ORDER BY [generate stable order]. Depending on the SQL dialect in use, there are different clauses that can be used to retrieve only the part of the result set you want (MySQL LIMIT clause, SQLServer TOP clause etc.)
If the number of documents is large the chance serving the same user the same document twice is neglibly small. Using the scheme described above you don't have to store any state information at all.
You may want to consider a nosql solution like Apache Cassandra. These seem to be ideally suited to your needs. There are many ways to design the algorithm you need in an environment where you can easily add new columns to a table (column family) on the fly, with excellent support for a very sparsely populated table.
edit: one of many possible solutions below:
create a CF(column family ie table) for each category (creating these on-the-fly is quite easy).
Add a row to each category CF for each document belonging to the category.
Whenever a user hits a document, you add a column with named and set it to true to the row. Obviously this table will be huge with millions of columns and probably quite sparsely populated, but no problem, reading this is still constant time.
Now finding a new document for a user in a category is simply a matter of selecting any result from select * where == null.
You should get constant time writes and reads, amazing scalability, etc if you can accept Cassandra's "eventually consistent" model (ie, it is not mission critical that a user never get a duplicate document)
I've solved similar in the past by indexing the relational database into a document oriented form using Apache Lucene. This was before the recent rise of NoSQL servers and is basically the same thing, but it's still a valid alternative approach.
You would create a Lucene Document for each of your texts with a textId (relational database id) field and multi valued categoryId and userId fields. Populate the categoryId field appropriately. When a user reads a text, add their id to the userId field. A simple query will return the set of documents with a given categoryId and without a given userId - pick one randomly and display it.
Store a users past X selections in a cookie or something.
Return the last selections to the server with the users new criteria
Randomly choose one of the texts satisfying the criteria until it is not a member of the last X selections of the user.
Return this choice of text and update the list of last X selections.
I would experiment to find the best value of X but I have in mind something like an X of say 16?
I am designing a table in Teradata with about 30 columns. These columns are going to need to store several time-interval-style values such as Daily, Monthly, Weekly, etc. It is bad design to store the actual string values in the table since this would be an attrocious repeat of data. Instead, what I want to do is create a primitive lookup table. This table would hold Daily, Monthly, Weekly and would use Teradata's identity column to derive the primary key. This primary key would then be stored in the table I am creating as foreign keys.
This would work fine for my application since all I need to know is the primitive key value as I populate my web form's dropdown lists. However, other applications we use will need to either run reports or receive this data through feeds. Therefore, a view will need to be created that joins this table out to the primitives table so that it can actually return Daily, Monthly, and Weekly.
My concern is performance. I've never created a table with such a large amount of foreign key fields and am fairly new to Teradata. Before I go on the long road of figuring this all out the hard way, I'd like any advice I can get on the best way to achieve my goal.
Edit: I suppose I should add that this lookup table would be a mishmash of unrelated primitives. It would contain group of values relating to time intervals as already mentioned above, but also time frames such as 24x7 and 8x5. The table would be designed like this:
ID Type Value
--- ------------ ------------
1 Interval Daily
2 Interval Monthly
3 Interval Weekly
4 TimeFrame 24x7
5 TimeFrame 8x5
Edit Part 2: Added a new tag to get more exposure to this question.
What you've done should be fine. Obviously, you'll need to run the actual queries and collect statistics where appropriate.
One thing I can recommend is to have an additional row in the lookup table like so:
ID Type Value
--- ------------ ------------
0 Unknown Unknown
Then in the main table, instead of having fields as null, you would give them a value of 0. This allows you to use inner joins instead of outer joins, which will help with performance.
I have designed my database in such a way that One of my table contains 52 columns. All the attributes are tightly associated with the primary key attribute, So there is no scope of further Normalization.
Please let me know if same kind of situation arises and you don't want to keep so many columns in a single table, what is the other option to do that.
It is not odd in any way to have 50 columns. ERP systems often have 100+ columns in some tables.
One thing you could look into is to ensure most columns got valid default values (null, today etc). That will simplify inserts.
Also ensure your code always specifies the columns (i.e no "select *"). Any kind of future optimization will include indexes with a subset of the columns.
One approach we used once, is that you split your table into two tables. Both of these tables get the primary key of the original table. In the first table, you put your most frequently used columns and in the second table you put the lesser used columns. Generally the first one should be smaller. You now can speed up things in the first table with various indices. In our design, we even had the first table running on memory engine (RAM), since we only had reading queries. If you need to get the combination of columns from table1 and table2 you need to join both tables with the primary key.
A table with fifty-two columns is not necessarily wrong. As others have pointed out many databases have such beasts. However I would not consider ERP systems as exemplars of good data design: in my experience they tend to be rather the opposite.
Anyway, moving on!
You say this:
"All the attributes are tightly associated with the primary key
attribute"
Which means that your table is in third normal form (or perhaps BCNF). That being the case it's not true that no further normalisation is possible. Perhaps you can go to fifth normal form?
Fifth normal form is about removing join dependencies. All your columns are dependent on the primary key but there may also be dependencies between columns: e.g, there are multiple values of COL42 associated with each value of COL23. Join dependencies means that when we add a new value of COL23 we end up inserting several records, one for each value of COL42. The Wikipedia article on 5NF has a good worked example.
I admit not many people go as far as 5NF. And it might well be that even with fifty-two columns you table is already in 5NF. But it's worth checking. Because if you can break out one or two subsidiary tables you'll have improved your data model and made your main table easier to work with.
Another option is the "item-result pair" (IRP) design over the "multi-column table" MCT design, especially if you'll be adding more columns from time to time.
MCT_TABLE
---------
KEY_col(s)
Col1
Col2
Col3
...
IRP_TABLE
---------
KEY_col(s)
ITEM
VALUE
select * from IRP_TABLE;
KEY_COL ITEM VALUE
------- ---- -----
1 NAME Joe
1 AGE 44
1 WGT 202
...
IRP is a bit harder to use, but much more flexible.
I've built very large systems using the IRP design and it can perform well even for massive data. In fact it kind of behaves like a column organized DB as you only pull in the rows you need (i.e. less I/O) rather that an entire wide row when you only need a few columns (i.e. more I/O).