Strange failure in test suite when removing code for caching the engine

[fsaad]

When changing CGPM_Metamodel to not use an engine cache and instead always load the engine from disk (which is what cache.patch does in the example session below), then test_cgpm.test_cgpm_smoke fails: a BQL query which specifies LIMIT 1 ends up returning 2 results. Strange!

$ git checkout 50af2c99446aa55df38c5a8e3c0cea2d53b816f2
$ ./check.sh -k test_cgpm_smoke tests/test_cgpm.py
==== 1 passed, 8 deselected in 2.03 seconds ========
$ cat cache.patch 
diff --git a/src/metamodels/cgpm_metamodel.py b/src/metamodels/cgpm_metamodel.py
index 6aa09b9..39ad955 100644
--- a/src/metamodels/cgpm_metamodel.py
+++ b/src/metamodels/cgpm_metamodel.py
@@ -696,8 +696,8 @@ class CGPM_Metamodel(IBayesDBMetamodel):
     def _engine(self, bdb, generator_id):
         # Probe the cache.
         cache = self._cache(bdb)
-        if cache is not None and generator_id in cache.engine:
-            return cache.engine[generator_id]
+        # if cache is not None and generator_id in cache.engine:
+        #     return cache.engine[generator_id]

         # Not cached.  Load the engine from the database.
         cursor = bdb.sql_execute('''
$ git apply cache.patch
$ ./check.sh -k test_cgpm_smoke tests/test_cgpm.py
bdb = <bayeslite.bayesdb.BayesDB object at 0x7f9411378e10>, gen = None, vars = ['output', 'cat', 'input']

    def cgpm_smoke_tests(bdb, gen, vars):
        modelledby = 'MODELLED BY %s' % (gen,) if gen else ''
        for var in vars:
            bdb.execute('''
                ESTIMATE PROBABILITY DENSITY OF %s = 1 WITHIN p %s
            ''' % (var, modelledby)).fetchall()
            bdb.execute('''
                SIMULATE %s FROM p %s LIMIT 1
            ''' % (var, modelledby)).fetchall()
            bdb.execute('''
                INFER %s FROM p %s LIMIT 1
            ''' % (var, modelledby)).fetchall()
            nvars = len(bdb.execute('''
                ESTIMATE * FROM VARIABLES OF p %(modelledby)s
                    ORDER BY PROBABILITY OF
                        (MUTUAL INFORMATION WITH %(var)s USING 1 SAMPLES > 0.1)
            ''' % {'var': var, 'modelledby': modelledby}).fetchall())
            if 0 < nvars:
                c = bdb.execute('''
                    SIMULATE p.(ESTIMATE * FROM VARIABLES OF p %(modelledby)s
                                    ORDER BY PROBABILITY OF
                                        (MUTUAL INFORMATION WITH %(var)s
                                            USING 1 SAMPLES > 0.1))
                        FROM p
                        LIMIT 1
                ''' % {'var': var, 'modelledby': modelledby}).fetchall()
>               assert len(c) == 1
E               assert 2 == 1
E                +  where 2 = len([('-1', 2.7210143958848385, 0.5285247869242387), ('1', 5.794596701327366, 4.578269369037486)])
==== 1 failed, 8 deselected in 2.03 seconds ========

@riastradh-probcomp You wrote the test in question -- any idea why it fails when removing the engine cache? Also, what does caching the engine have to do with LIMIT 1 returning 2 rows?

[riastradh-probcomp]

Dunno, the first thing I would do is examine the SQL trace to see what SQL queries are actually getting executed.

[fsaad]

Turned out to be a massively stochastic bug which is unrelated to engine caching.