Optional attributes fixes #88
Conversation
Conversions targeting types which are or include objects with optional attributes always simplify those types, removing the optional attributes from the type, and inserting null values for any that are missing. Objects with optional attributes are not intended to be used to instantiate values. One accidental exception to this was the null value of dynamic type. These values are passed through by the getConversion wrapper, assigning the desired type. This conflicts with the type simplification rules for non-null values, and can result in a panic when a collection type contains null values. This commit fixes this at the wrapper level by performing the same type simplification when converting null into an object type.
To correctly unify deep nested types, we need to unify all collection types (maps, lists, sets), not just maps. This ensures that all cousin types are unified correctly. This problem was first addressed for maps in zclconf#47 a year ago, but we didn't realize at the time that the same bug could manifest in other collection types.
Codecov Report
@@ Coverage Diff @@
## main #88 +/- ##
==========================================
+ Coverage 70.67% 70.69% +0.02%
==========================================
Files 79 79
Lines 6547 6580 +33
==========================================
+ Hits 4627 4652 +25
- Misses 1476 1483 +7
- Partials 444 445 +1
Continue to review full report at Codecov.
|
There was a problem hiding this comment.
Thanks for digging into these, @alisdair!
The overall approach here looks good to me, though please see my inline comment below about a trickier edge that I think this isn't covering.
cty/convert/conversion.go
Outdated
| // conversion function which drops the optional attributes from the | ||
| // type. Null pass through values must do the same to ensure that | ||
| // homogeneous collections have a single element type. | ||
| if out.IsObjectType() && len(out.OptionalAttributes()) > 0 { |
There was a problem hiding this comment.
Seeing this check made me wonder if the shallow check here would be sufficient to deal with a case where we have a type constraint that only indirectly contains an object type with optional attributes. In order to try to test that I added the following test case in public_test.go, adapting one of the ones you added in this PR:
{
Value: cty.TupleVal([]cty.Value{
cty.NullVal(cty.DynamicPseudoType),
cty.ObjectVal(map[string]cty.Value{
"foo": cty.NullVal(cty.DynamicPseudoType),
}),
}),
Type: cty.List(cty.Object(
map[string]cty.Type{
"foo": cty.ObjectWithOptionalAttrs(
map[string]cty.Type{
"bar": cty.String,
},
[]string{"bar"},
),
},
)),
Want: cty.ListVal([]cty.Value{
cty.NullVal(cty.Object(map[string]cty.Type{
"foo": cty.Object(map[string]cty.Type{
"bar": cty.String,
}),
})),
cty.ObjectVal(map[string]cty.Value{
"foo": cty.NullVal(cty.Object(map[string]cty.Type{
"bar": cty.String,
})),
}),
}),
},Unfortunately this seems to fall into the same panic trap as the original upstream report 😖
panic: inconsistent list element types (cty.Object(map[string]cty.Type{"foo":cty.ObjectWithOptionalAttrs(map[string]cty.Type{"bar":cty.String}, []string{"bar"})}) then cty.Object(map[string]cty.Type{"foo":cty.Object(map[string]cty.Type{"bar":cty.String})})) [recovered]
panic: inconsistent list element types (cty.Object(map[string]cty.Type{"foo":cty.ObjectWithOptionalAttrs(map[string]cty.Type{"bar":cty.String}, []string{"bar"})}) then cty.Object(map[string]cty.Type{"foo":cty.Object(map[string]cty.Type{"bar":cty.String})}))
goroutine 156 [running]:
testing.tRunner.func1.2(0x5ad520, 0xc0001be020)
/home/me/.goenv/versions/1.16.0/src/testing/testing.go:1144 +0x332
testing.tRunner.func1(0xc0001abb00)
/home/me/.goenv/versions/1.16.0/src/testing/testing.go:1147 +0x4b6
panic(0x5ad520, 0xc0001be020)
/home/me/.goenv/versions/1.16.0/src/runtime/panic.go:965 +0x1b9
github.com/zclconf/go-cty/cty.ListVal(0xc000160f80, 0x2, 0x2, 0xc0001483c0, 0xc00017ff00, 0x1, 0x1)
/home/me/Devel/go-cty/cty/value_init.go:166 +0x57e
github.com/zclconf/go-cty/cty/convert.conversionCollectionToList.func1(0x6b97d0, 0xc000133540, 0x5a40e0, 0xc0000bd2f0, 0x0, 0x0, 0x0, 0x557e14, 0xc00009ad68, 0x40da5b, ...)
/home/me/Devel/go-cty/cty/convert/conversion_collection.go:54 +0x793
github.com/zclconf/go-cty/cty/convert.getConversion.func1(0x6b97d0, 0xc000133540, 0x5a40e0, 0xc0000bd2f0, 0x0, 0x0, 0x0, 0xc0001a3ac0, 0xc00017fef0, 0x6b97d0, ...)
/home/me/Devel/go-cty/cty/convert/conversion.go:56 +0x52c
github.com/zclconf/go-cty/cty/convert.retConversion.func1(0x6b97d0, 0xc000133540, 0x5a40e0, 0xc0000bd2f0, 0xc00017fef0, 0x4d33a4, 0x76c940, 0x6d17d2, 0xf, 0x6e0f68)
/home/me/Devel/go-cty/cty/convert/conversion.go:198 +0x6b
github.com/zclconf/go-cty/cty/convert.Convert(0x6b97d0, 0xc000133540, 0x5a40e0, 0xc0000bd2f0, 0x6b97d0, 0xc000133570, 0xf, 0x6e0f68, 0x38, 0x439, ...)
/home/me/Devel/go-cty/cty/convert/public.go:51 +0x1c5
github.com/zclconf/go-cty/cty/convert.TestConvert.func1(0xc0001abb00)
/home/me/Devel/go-cty/cty/convert/public_test.go:986 +0x97
testing.tRunner(0xc0001abb00, 0xc00017fe80)
/home/me/.goenv/versions/1.16.0/src/testing/testing.go:1194 +0xef
created by testing.(*T).Run
/home/me/.goenv/versions/1.16.0/src/testing/testing.go:1239 +0x2b3
I think unfortunately getting correct behavior here is going to require a recursive walk of the type constraint for any null value (and possibly also the same for ths !in.IsKnown() case above?) to concrete-ize any of the object types with optional attributes inside. The method cty.Type.HasDynamicTypes shows a pattern for implementing a recursive walk of that sort, so I guess my first idea is to add cty.Type.WithoutOptionalAttributesDeep that will return a type with all of the optional attributes inside removed. That will unfortunately be a bit expensive for new memory allocation, but I think in a way that will have similar impact to the "normal" (non-null, known) type conversion codepaths.
There was a problem hiding this comment.
@apparentlymart I'm unable to reproduce this panic when both of the commits on this branch are in place. The extra unification step introduced for lists (and sets) handles the disjoint types of the null and object tuple members, and the result is that the test passes. If I revert the unification commit, the panic occurs.
With unification in place, I can't see any situation where the WithoutOptionalAttributesDeep function is needed. I'm happy to add a commit to add and use it, but I'm not sure it's necessary until I can find a test case which fails without it.
Am I missing something, or did you test the panic against the first commit only? Any suggestions for an extended test which would fail in this way?
There was a problem hiding this comment.
Hi @alisdair,
Curiously, I went back to my dev environment where I'd previously tested this and re-ran tests without any modifications and it didn't panic for me anymore either, so I have to assume I'd somehow got a stale result out of the test suite, although I have no concrete explanation as to why that would be.
Your explanation of how the unification step should get expected result makes sense to me now I re-read the changeset with that idea in mind, and I can no longer reproduce this panic, so instead of trying to explain this oddity I'm going to just assume I made a human error while I was testing and inadvertently tested something other than what I thought I was testing.
With that said, I made some adjustments to the test to take type unification out of the process and focus only on the type conversion aspect, with this new test case:
{
Value: cty.NullVal(cty.DynamicPseudoType),
Type: cty.Object(
map[string]cty.Type{
"foo": cty.ObjectWithOptionalAttrs(
map[string]cty.Type{
"bar": cty.String,
},
[]string{"bar"},
),
},
),
Want: cty.NullVal(cty.Object(map[string]cty.Type{
"foo": cty.Object(map[string]cty.Type{
"bar": cty.String,
}),
})),
},This failed for me because the resulting null value had a type still containing optional attributes:
--- FAIL: TestConvert/cty.NullVal(cty.DynamicPseudoType)_to_cty.Object(map[string]cty.Type{"foo":cty.ObjectWithOptionalAttrs(map[string]cty.Type{"bar":cty.String},_[]string{"bar"})}) (0.00s)
public_test.go:987: wrong result
value: cty.NullVal(cty.DynamicPseudoType)
type: cty.Object(map[string]cty.Type{"foo":cty.ObjectWithOptionalAttrs(map[string]cty.Type{"bar":cty.String}, []string{"bar"})})
got: cty.NullVal(cty.Object(map[string]cty.Type{"foo":cty.ObjectWithOptionalAttrs(map[string]cty.Type{"bar":cty.String}, []string{"bar"})}))
want: cty.NullVal(cty.Object(map[string]cty.Type{"foo":cty.Object(map[string]cty.Type{"bar":cty.String})}))
Of course, we could decide that the Want value I wrote here is what's wrong, rather than the actual result, but given that in other cases we've decided that the result should have all of the optional attribute annotations removed, I wanted to raise this to see what you think about the inconsistency of this result.
There was a problem hiding this comment.
Thanks for the extra clarification! I think the most recent commit addresses this now, for both unknown and null values.
When converting unknown and null values to a given type, we want to avoid constructing instances of types which include optional object attributes, even deeply nested. To do this we add a new method, cty.Type.WithoutOptionalAttributesDeep, which recursively processes its receiver to return a type which has no instances of objects with optional attributes.
Two commits addressing issues related to objects with optional attributes. These commits address these downstream bugs in Terraform:
Fix for null objects with optional attrs
Conversions targeting types which are or include objects with optional attributes always simplify those types, removing the optional attributes from the type, and inserting null values for any that are missing. Objects with optional attributes are not intended to be used to instantiate values.
One accidental exception to this was the null value of dynamic type. These values are passed through by the getConversion wrapper, assigning the desired type. This conflicts with the type simplification rules for non-null values, and can result in a panic when a collection type contains null values.
This commit fixes this at the wrapper level by performing the same type simplification when converting null into an object type.
Fix unification of collection values
To correctly unify deep nested types, we need to unify all collection types (maps, lists, sets), not just maps. This ensures that all cousin types are unified correctly.
This problem was first addressed for maps in #47 a year ago, but we didn't realize at the time that the same bug could manifest in other collection types.