how to load data faster with talend and sql server

Question

I use Talend to load data into a sql-server database.

It appears that the weakest point of my job is not the dataprocessing, but the effective load in my database, which is not faster than 17 rows/sec.

The funny point is that I can launch 5 jobs in the same time, and they'll all load at 17rows/sec .

What could explain this slowness and how could I improve the speed?

Thanks

New informations:

The transfer speed between my desktop and the server is about 1MByte

My job commits every 10 000

I use sql server 2008 R2

And the schema I use for my jobs is like this:

Answer 1

Database INSERT OR UPDATE methods are incredibly costly as the database cannot batch all of the commits to do all at once and must do them line by line (ACID transactions force this because if it attempted to do an insert and then failed then all of the other records in this commit would also fail).

Instead, for large bulk operations it is always best to predetermine whether a record would be inserted or updated before passing the commit to the database and then sending 2 transactions to the database.

A typical job that needed this functionality would assemble the data that is to be INSERT OR UPDATEd and then query the database table for the existing primary keys. If the primary key already exists then you can send this as an UPDATE, otherwise it is an INSERT. The logic for this can be easily done in a tMap component.

In this job we have some data that we wish to INSERT OR UPDATE into a database table that contains some pre-existing data:

And we wish to add the following data to it:

The job works by throwing the new data into a tHashOutput component so it can be used multiple times in the same job (it simply puts it to memory or in large instances can cache it to the disk).

Following on from this one lot of data is read out of a tHashInput component and directly into a tMap. Another tHashInput component is utilised to run a parameterised query against the table:

You may find this guide to Talend and parameterised queries useful. From here the returned records (so only the ones inside the database already) are used as a lookup to the tMap.

This is then configured as an INNER JOIN to find the records that need to be UPDATED with the rejects from the INNER JOIN to be inserted:

These outputs then just flow to separate tMySQLOutput components to UPDATE or INSERT as necessary. And finally when the main subjob is complete we commit the changes.

Answer 2

I think that @ydaetskcoR 's answer is perfect on a teorical point of view (divide rows that need Insert from those to Update) and gives you a working ETL solution useful for small dataset (some thousands rows).

Performing the lookup to be able to decide wheter a row has to be updated or not is costly in ETL as all the data is going back and forth between the Talend machine and the DB server.

When you get to some hundred of thousands o even millions of records you have to pass from ETL to ELT: you just load your data to some temp (staging) table as suggested from @Balazs Gunics and then you use SQL to manipulate it.

In this case after loading your data (only INSERT = fast, even faster using BULK LOAD components) you will issue a LEFT OUTER JOIN between the temp table and the destination one to divide the rows that are already there (need update) and the others.

This query will give you the rows you need to insert:

SELECT staging.* FROM staging
LEFT OUTER JOIN destination ON (destination.PK = staging.PK)
WHERE destination.PK IS NULL

This other one the rows you need to update:

SELECT staging.* FROM staging
LEFT OUTER JOIN destination ON (destination.PK = staging.PK)
WHERE destination.PK IS   NOT    NULL

This will be orders of magnitude faster than ETL, BUT you will need to use SQL to operate on your data, while in ETL you can use Java as ALL the data is taken to the Talend server, so often is common a first step on the local machine to pre-process the data in java (to clean and validate it) and then fire it up on the DB where you use join to load it in the right way.

Here are the ELT JOB screen shots.

Answer 3

I've found where this performance problem come form.

I do an INSERT OR UPDATE, if I replace it with a simple INSERT, the speed goes up to 4000 rows/s.

Does it seem like an acceptable pace?

Anyway, I need my INSERT OR UPDATE so, I guess I'm stuck.

Answer 4

Based on your note that inserts are an order of magnitude faster than updates (4000 vs 17/sec) - It looks like you need to look at your DB indexes. Adding an index that matches your update parameters could speedup your updates significantly. Of course, this index may slow your inserts a bit.

You can also look at the query execution plan for your update query to see if it is using any indexes. How do I obtain a Query Execution Plan?

Answer 5

You should do a staging table, where you insert the rows.

Based on this staging table you do a DELETE query with t*SQLrow.

DELETE FROM target_table
WHERE target_table.id IN (SELECT id FROM staging_table);

So the rows you wanted to update are no longer exists.

INSERT INTO target_table 
SELECT * FROM staging_table;

This will move all the new/modified rows.

Answer 6

I was having the same issue loading data into a DB2 server. I too had the commit set at 10000 but once I selected the option to batch(on the same component options screen) performance dramatically improved. When I moved the commit and batch to 20000 the job went from 5 hours to under 2 minutes.

Answer 7

I had the same problem and solved it by defining an index on target table.

Usually, the target table has an id field which is its primary key and hence indexed. So, all sort of joins with it would work just fine. But the update from a flat file is done by some data fields, so each update statement have to make full table scan.

The above also explains why it works fast with INSERT and becomes slow with INSERT OR UPDATE

Answer 8

Recommend a couple simple things:

1. Where possible / reasonable, change from ETL to ELT.
2. Set up a CDC process and only handle the changes. Depending on the database and needs, this can be handled (a) directly on the database, (b) through automated Talend functionality (need a subscription), (c) manually via SQL (full outer join) and a custom Java function that generates an MD5 hash, or (d) manually via SQL (full outer join) and the tAddCRCRow component.
3. Where possible, load multiple tables concurrently.
4. Where possible, use bulk loading for tables.
5. Sometimes, a clear and load is acceptable as an approach and faster than checking for updates.