Aggregator Limit

[Raftsman]

I receive the following error during aggregation;

Aggregator_37,0: Failure during execution of operator logic.
Aggregator_37,0: Input 0 consumed 9614492 records.
Aggregator_37,0: Output 0 produced 9614492 records.
Aggregator_37,0: Fatal Error: pipe write failed: Broken pipe
Aggregator_37,0: Failure during execution of operator logic.
Aggregator_37,0: Input 0 consumed 9614046 records.
Aggregator_37,0: Output 0 produced 6683230 records.
Aggregator_37,0: Fatal Error: sendReadAck(): write failed on node SASHQOKWSDA Broken pipe
node_node2: Player 18 terminated unexpectedly.

I assume it's a space issue as the program worked fine with a 5 million row output. I ran it multiple time and it keeps aborting around the same amount of records. Is there a limit that I should change.

Thanks,

[ArndW]

Have you monitored your scratch space while the job was running? Perhaps the broken pipe was caused by space issues.

[dsusr]

Are you doing some pre-sorting before aggregation ?

Definitely its an space/memory issue. Temporary fix can be to run the job as sequential or make aggregator stage as sequential.

[wesd]

The aggregator is known to choke under large volumes of data. Check space and rearchitect the job if necessary.

[chulett]

If it's anything like the Server Aggregator stage, you can help it out tremendously by presorting your data in a manner than support the grouping being done. This should, in essence, remove any 'large volumes' issues.

[sud]

I think in the aggregate method you are using hash, use "sort" instead.

[Raftsman]

Follow-up:

We have two servers. DEV and PROD If I run the same job on each server, I get two different results. On the DEV server, the job completes normally. On the PROD server, the job aborts. The only DS difference is a hot patch IBM offered to fix another problem. I have a suspicion that the patch has caused the problem. I will keep you posted.

I will also try some of your suggestions.

Thanks

[Raftsman]

Well after 5 weeks of grueling pain, I think we have solved this issue. Today June 18th, IBM will be releasing a patch regarding the MKS Toolkit which solves many broken pipe problems. If you are a Windows client and receive many broken pipe issues, I suggest getting the patch from IBM.

[abc123]

Raftsman, do you have a patch number that you can give me?

[chulett]

Five weeks? Looks more like a year and a half to me.