QC Lockstep tests: tweak test case generation

[dcoutts]

This is an attempt to get the QLS tests to get more coverage of the duplicate table feature, and more generally to have more interesting coverage.

In particular, we move several of the boring special cases out of the QLS test and into unit tests. This is for things like operating on closed table handles, or opening non-existent snapshots. These things of course must be covered, but they don't have interesting interactions so they don't need to be covered in complex situations. This lets the limited length of action sequences to be used for more interesting things.

See quickcheck-dynamic#85 for further details on why action sequences are so short.

We also change how actions are generated to have fewer tables, and thus more actions per table. We limit the number of tables at once to 5.

Initial results (from 28c4890):

      Interleaved actions on table duplicates (270 in total):
      88.1% DupTableActionLog "1 <= n < 2"
       6.3% DupTableActionLog "2 <= n < 4"
       3.3% DupTableActionLog "4 <= n < 8"
       1.9% DupTableActionLog "8 <= n < 16"
       0.4% DupTableActionLog "16 <= n < 32"

and now we get:

      Interleaved actions on table duplicates (125 in total):
      55.2% DupTableActionLog "1 <= n < 2"
      12.8% DupTableActionLog "4 <= n < 8"
      11.2% DupTableActionLog "2 <= n < 4"
      11.2% DupTableActionLog "8 <= n < 16"
       7.2% DupTableActionLog "16 <= n < 32"
       2.4% DupTableActionLog "32 <= n < 64"

So this is considerably improved, with 45% of runs now having non-trivial interleavings rather than 12%. And where previously we had only 2.3% of original tables with interleavings of length 8 or more, we now have over 20%.

One important caveat is that the change in the key and value types mean the tests run longer. Summary of running times from my system:

    propLockstep_ModelIOImpl:           OK (11.15s)
    propLockstep_RealImpl_RealFS_IO:    OK (24.24s)
    propLockstep_RealImpl_MockFS_IO:    OK (77.06s)
    propLockstep_RealImpl_MockFS_IOSim: OK (111.07s)

[jorisdral]

Nice!

Tests are timing out because CI currently only allows tests to run for around 5 minutes. These lines will have to be tweaked:

lsm-tree/.github/workflows/haskell.yml

Lines 123 to 133 in 3f5bda8

	# https://docs.github.com/en/actions/using-workflows/workflow-commands-for-github-actions#setting-an-environment-variable
	- name: Set test timeout (Unix)
	if: ${{ startsWith(matrix.os, 'ubuntu') || startsWith(matrix.os, 'macOS') }}
	run: |
	echo "TASTY_TIMEOUT=5m" >> "$GITHUB_ENV"
	# https://github.com/actions/runner/issues/2281#issuecomment-1326748709
	- name: Set test timeout (Windows)
	if: ${{ startsWith(matrix.os, 'windows') }}
	run: |
	echo "TASTY_TIMEOUT=5m" >> "$env:GITHUB_ENV"

[dcoutts]

In the first version of this PR the running times on my machine were:

    propLockstep_ModelIOImpl:           OK (17.80s)
    propLockstep_RealImpl_RealFS_IO:    OK (86.42s)
    propLockstep_RealImpl_MockFS_IO:    OK (251.20s)
    propLockstep_RealImpl_MockFS_IOSim: OK (284.86s)
Test.Database.LSMTree.Normal.StateMachine.DL
    prop_example:                       OK (455.09s)

But with a couple changes they're down to this:

    propLockstep_ModelIOImpl:           OK (11.15s)
    propLockstep_RealImpl_RealFS_IO:    OK (24.24s)
    propLockstep_RealImpl_MockFS_IO:    OK (77.06s)
    propLockstep_RealImpl_MockFS_IOSim: OK (111.07s)

Those changes are:

• Use SerialisedBlob rather than LargeRawBytes for the Blob type.
• Disable the StateMachine.DL.prop_example test

In particular, the latter was not justifying its long running time with any crucial coverage.

[dcoutts]

Oh and so hopefully this will mean we do not need to adjust the CI testsuite timeouts. 🤞

[jorisdral]

LGTM! There is just the hlint failure to fix, but then we can merge