units.parse has inconsistent rounding errors for values of 0.5

[tmos22]

Short description

units.parse has rounding errors when calculating 0.5 during conversion with the milli- suffix (lowercase m) and Kilo- (k/K). All other prefixes seem to calculate properly.

System Details:

• OPA version: >= v0.41.0 (checked against both 0.41.0 and 0.42.0)
• OS: RHEL 7.9, UEK 5.4.17
• Go Version: go1.16.4 linux/amd64
• Built locally with the following command/flags

 go build -o opa -ldflags "\
		-X github.com/open-policy-agent/opa/internal/report.ExternalServiceURL='' \
		-X github.com/open-policy-agent/opa/version.Version=local-unofficial \
		-X github.com/open-policy-agent/opa/version.Vcs=`./build/get-build-commit.sh` \
		-X github.com/open-policy-agent/opa/version.Timestamp=`./build/get-build-timestamp.sh` \
		-X github.com/open-policy-agent/opa/version.Hostname=`./build/get-build-hostname.sh`"

Steps To Reproduce

1. Execute opa run
2. Run units.parse("500m")
3. Receive rounding error of 0.5000000000000000104
4. Run units.parse("0.0005K")
5. Receive rounding error of 0.49999999999999999997

> units.parse("0.0000000000000000005E")
0.5
> units.parse("0.0000000000000005P")
0.5
> units.parse("0.0000000000005T")
0.5
> units.parse("0.0000000005G")
0.5
> units.parse("0.0000005M")
0.5
> units.parse("0.0005K")
0.49999999999999999997
> units.parse("0.5")
0.5
> units.parse("500m")
0.5000000000000000104

# Testing equality
> units.parse("0.0000000000000000005E") == 0.5
true
> units.parse("0.0000000000000005P") == 0.5
true
> units.parse("0.0000000000005T") == 0.5
true
> units.parse("0.0000000005G") == 0.5
true
> units.parse("0.0000005M") == 0.5
true
> units.parse("0.0005K") == 0.5
false
> units.parse("0.5") == 0.5
true
> units.parse("500m") == 0.5
false

Expected behavior

All units that should be parsed to the same value should equal the same value (all 0.5s should be equal)

> units.parse("500m")
0.5
> units.parse("0.0005K")
0.5

[philipaconrad]

@tmos22 Thanks for reporting this! I was able to reproduce the issue locally from your test cases, and I have a pretty good idea of how to fix the issue. Internally, it's caused by floating-point rounding issues, even though OPA uses big.Float internally for the units parsing builtin.

The solution is probably going to involve normalizing the path that incoming units quantities go through, so that they experience identical rounding behavior, regardless of whether we're dealing with milli or Kilo units.