5

I have 2 type of tables to populate the data-warehouse with every day, lookup tables or configuration tables with few 100s records, and thats easy where i just truncate and refill the table.

but for transaction tables, that have many records, i usually increment, that is i run the ETL daily to add yesterdays records.

i have 2 problems that i face always

  1. when the job fails for any reason (i lose that Days transactions)
  2. when for any reason the job run twice or i run it twice (i get duplicates)

now i am trying to design a way where i over come these 2 problems as well as am trying to develop the ETL in such a way that it can auto fix it self incase any of these events occur.

i want it to check if there are missing days and run run the ETL for that day, and check if there are duplicates and delete them.

below are ways i though of 1. i take in the last 5 days regardless, every day the ETL runs, deletes the last 5 days and refill. 2. i check the destination tables if they have missing dates in the last month and then i query the source with the missing days.

keeping in mind that the source is a huge table in a production environment that i have to optimize my query to the maximum when requesting from it.

thanks

2
  • 3
    What is "huge"?
    – Philᵀᴹ
    Sep 1, 2013 at 7:47
  • i mean the source table has millions of records and many transactions per day, am talking about a banks ATM and Internet banking transactions
    – AmmarR
    Sep 1, 2013 at 7:50

4 Answers 4

2

Do the transactions have an audit timestamp ? It has to be one that goes up only (no late-arriving facts. insert/update audit timestamps are good for this)
If so you could use this to define a range to extract. It's a common technique for this type of thing:

  • For each extraction, at the start, determine the range you want to extract (let's call the minimum timestamp min_ts and the maximum timestamp max_ts)
  • At the start of the extraction, put 1 line in a separate table (let's call it extraction_log and also give it a PK) with fields: extraction_id = a unique key, min_ts, max_ts, status = 'Starting Extraction'.
  • Use min_ts & max_ts to extract the data, either in 1 go (select * from where ts > min_ts and ts <= max_ts) or in chunks if needed.
  • Ad the end of a successful extraction, update the line and set status to 'Finished OK'

how to determine min_ts and max_ts ?

  • You could take min_ts from the extraction_log, using the last successful max_ts. select max(max_ts) from extraction_log where status = 'Finished OK'

  • You could take max_ts from the source db at the start of your extraction. select max(audit_ts) from source_table

There are alternatives here. If you're extracting these into a temporary staging table (a best practice), you could also take too much, ie the last 5 days and deal with duplicates later when upserting the entries in your ODS. For max_ts, if you're absolutely certain that the clocks between your dwh and the source are and will remain in sync (this is a very dangerous assumption - not recommended), you could even use sysdate()

Technically, you can get away with less effort. You don't really need the statuses, or a extraction_log table that keeps track of each batch. But I have found having it like this helps a lot with debugging and troubleshooting later on. Additionally, if you need a routine to remove entries from a load that failed halfway, to find gaps for ranges of past extractions, and so on, the extraction_log will help. You might even want to include the extraction_id as an additional column in your ODS.

some further thoughts

  • if you don't have a good timestamp candidate, a technical key in the source system with the same properties (must go up only, no late-arriving facts) should be fine too.

  • if the timestamps are generated by the source application and there is a risk that 2 transactions with the exact same timestamp are not inserted at the same time (quite common), then it's safer to use a max_ts that is slightly in the past (select max(max_ts) - 5 minutes from source_table.)

  • regarding failed loads. if you're only concerned with cleaning up entries of the last failed loads, you can add it as a first step in your ETL flow. (delete from dwh_table where ts > min_ts) This will remove the entries from any failures (if any) after the last successful extraction. It won't deal with failures between earlier successful extractions.

1
  • +1. I've used these techniques successfully on a small-to-medium-sized (10s of millions of records) warehouse. Note that you can also cap your loads; e.g., select the first X records ordered by the ROWVERSION field, so you know your ETL will never attempt to move more than it can handle. If the system falls behind for any reason, it's better to catch up gradually than to try to import more than it can handle and fail when your log file fills up. Apr 10, 2015 at 15:33
1

Could you add an additional table to the data warehouse used for maintaining your imported transaction tables?

The maintenance table could be called something like:

ImportJob
    ImportID (primary key)
    TransactionDate

Add the ImportID column to the transaction tables and set this when loading data.

The import table would have one record for each day's worth of transactions being imported. A unique constraint on TransactionDate would prevent reloading data twice. You'd also have a fast way to delete out transactions with atomicity should the import fail. You can query the production tables for a day's data at a time and fill in any missing days.

1

It sounds like you're able to treat each day's records as a discrete unit. I'd try:

  1. Execute SQL Task or Data Flow - Get a list of dates from the DW for which there are no transactions, and shove the result into a recordset variable.
  2. For Each Container - Loop through the recordset and populate a date variable
  3. Data Flow - Build the source SQL in a variable with where transactiondate >= @missingdate and transactiondate < dateadd(day,1,@missingdate)

So if the job runs after having missed a day or two, it will query production separately for each day, but it shouldn't waste any resources pulling records that the DW already has. If the job is run multiple times in one day, it shouldn't even end up running a query against production, unless there's a day that didn't have any transactions (e.g. a holiday).

You may want to limit the date ranges for #1 above to the past 5 days, 30 days, or whatever makes sense.

The weakness in this approach is that the date list cannot contain the current or future dates, and that there is no provision for updates or new records for a date that has already been copied into the DW, but for your logged transactions, it sounds like that may be acceptable.

1

Alright this is going to be a sort of basic version of what exactly you need to first is going to be to make a control table. Essentially this will control the ETL you are doing. In this table one of the things you will want to store the last time the process ran successfully. Now when you're doing your ETL you can call to this table to find the date range you are looking for essentially all rows > LastSuccesfullRun

The next thing you want to do to prevent against duplicates is also fairly simple. First take your data flow task that has now been filtered by date range according to the last successfull run. Then use a lookup component to compare against the rowset you currently have in the table. On MATCH do nothing on NO Match insert. Attached below is a demo for this.

OLESource to Lookup

LookUp Table

No match

I would love to go into more detail but I don't know enough about your ETL process to give more advice.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.