Yossi Dahan [BizTalk]


Sunday, July 26, 2009

Parallel shape behaviour in BizTalk 2004 and >2006

‘Shiri’ had posted this question in the newsgroup -…

…After open the orchestration
debugger both at BT 2004 and BT 2006 we've recognized a different behaviour:
at 2004 executes through all the branches first shape and then back to the
first branch at parallel and executes the second ahape, the second branch
and etc. at 2006 executes the first branch - first and second shape, then
goes to the second branch and executes the first and second and etc.
It seems that 2004 works more like multithreaded then 2006….

I vaguely remember reading/discussing this difference in the past, but – unfortunately - I can't remember the exact details; I've floated this question around again, and I believe (but treat this with care – this may well be at least somewhat inaccurate) that there has indeed been a change in the behaviour of the parallel shape in BizTalk 2006, here’s some background -

The parallel shape was never intended to provide ‘true’ parallelism, and Microsoft has been fairly clear from the start that BizTalk will not process each branch on it’s own thread (which would have been required).

Darren Jefford explains very well the intention, and the expected behaviour, of the parallel shape in his book Professional BizTalk 2006, which I highly recommend; if you want a quick peek you can read the relevant piece here, the key point he makes is that when thinking about the parallel shape, you need to wear your business analyst hat, and not the developer hat – the parallel shape effectively “says” – hey - I’ll run this code (=branch), but if I reach a point where I’m sitting idle waiting for something to happen (receive, delay or listen shapes) I will go and run that other code (=branch) in the meantime; when I’m through with that (or reached a waiting point again) I will check if I’m ready to process the rest of the first bit of code…and so on…

This is not quite your techie run-things-in-parallel-on-multiple-threads approach, but – from my experience – it is more than enough (if you wanted to run things completely in parallel you could use BizTalk’s pub/sub, which would allow you to potentially get a lot more than just one thread – you might end up on a different machine altogether, for a price :-))

So – indeed – the BizTalk 2006’s (and subsequent versions) parallel shape behaves exactly as I understand it should do (and as is described in Darren’s book) and is consistent with Shiri’s observation - if you have three branches, and neither have a blocking shape – the left most branch will be executed completely, then the next one to its right and then the right most branch; however –and that’s a very important point to remember - Microsoft does not, to the best of my knowledge, guarantee any order of processing between the branches, and this might change in future versions (as indeed it has been seen from 2004 to 2006), so all you can assume is that all the branches can theoretically run in any arbitrary order or indeed in parallel.

Back to the 2006 behaviour - if, however, you had a receive shape as the second shape in the left most branch, when BizTalk would hit this shape it would move on to execute the second branch while it’s waiting for the message to be received; it would come back to the first branch at the earliest point once two things had happened – 1. the message it was waiting for was delivered and 2. it had reached a point in the currently executing branch in which it can stop and re-enter the first branch; this would be a receive shape, delay shape, listen shape or the end of the branch.

So – if that’s the 2006 behaviour, was the behaviour in 2004 different? yes, I believe it was – in 2004 the engine was, some would say, trying to be too clever - BizTalk 2004 would try to run branches on multiple thread if it can; where “if it can” depends on several factors, not the least the state of the thread pool at the time of evaluation; if it managed to do so, you would get code running truly in parallel, as Shiri observed, but there are no guarantees that this would be the case; in that sense BizTalk 2004 is less predictable than later versions of BizTalk, which is exactly the problem with this approach, and - considering that this was never the intention to begin with – I can fully understand the decision to simplify the model in BizTalk 2006.

Labels: ,


Post a Comment

<< Home