dry run slow for complex process graph

[jdries]

For sufficiently complex process graphs, the dry run becomes the bottleneck.
The dry run does not yet perform caching, because this was deemed safer, especially to determine correct 'source constraints' in complex graphs.

I'll put the actual graph in a PR.

The profiler also shows some hotspots:

In addition, I noticed that the test takes a large amount of memory, I had to limit it on process level to stop my IDE from crashing. Removing some debug logging did help, so it could be related to pytest feature of capturing logs (in memory).

If I disable logging, the pure process graph evaluation does run in under 1 minute. Now the 'get_source_constraints' seems to be the problem.

[jdries]

PR merged, will close the issue now, but to be confirmed after rollout.
Introducing caching in dry run may also help, but that has a bigger chance of introducing unexpected side effects.

In unit tests, the first run without optimizations was hanging for 2+ hours until I aborted it.
After first optimization it was down to 27 minutes, but then we disabled the test. (So there was significant improvement, to be seen how it works out in the live environments.)