PEREL-5643: support Gi/Mi strings in autotune_limits mem/disk clamping

[matt-ullmer]

Summary

• Add _resource_value_to_mib() to normalize resource values (int/float MiB or K8s strings like "4Gi"/"512Mi") to a common unit before comparing
• Update _clamp_recommendations() to use it so Cassandra clusters can set mem/disk floors via autotune_limits without TypeError on string comparisons
• Fix pre-existing log bug: max-clamping branch was logging min_value instead of max_value

Cassandra autotune recommendations use Gi strings for mem/disk. Direct </> comparisons against integer limits would raise TypeError. This is the paasta-side change; a follow-up yelpsoa-configs PR will update the cassandra_k8s/validate.py schema from integer to string pattern once this is deployed.

Jira: https://jira.yelpcorp.com/browse/PEREL-5643

Test plan

• pytest tests/test_config_utils.py — all 37 tests pass including new parametrized test_resource_value_to_mib cases and test_auto_config_updater_merge_recommendations_limits_gi_strings
• pre-commit clean on changed files

[nemacysts]

hmmm, the naming here is gonna be a bit confusing - this is used for all 3 resource types (cpu, mem, disk), but only applies to mem/disk

...not sure if we wanna branch of the type or if we want to just have a comment here explaining the weirdness

[nemacysts]

@matt-ullmer i think this is the only outstanding comment now - i'm not really sure what's best though :p

i think another potential option might be something like

for limit_type, limits in config.get("autotune_limits", {}).items():
            log.debug(f"Processing {limit_type} autotune limits...")
            min_value = _normalized_value(limits.get("min"), limit_type)
            max_value = _normalized_value(limits.get("max"), limit_type)
            unclamped_resource_value = _normalized_value(clamped_recomendation.get(limit_type), limit_type)

where _normalized_value is something like:

NORMALIZABLE_VALUE_TYPES = {"mem", "disk", ...}

...

def _normalized_value(value: int | float | str, value_type: str) -> float: # XXX: maybe an enum/Literal/etc for value_type would work?
    """Normalize a resource value to MiB for comparison. For numeric types
    (e.g. cpus as float), this is a passthrough cast to float."""

    if isinstance(value, (int, float)):
        return float(value)
    elif value_type in NORMALIZABLE_VALUE_TYPES and isinstance(value, str):
        for suffix, factor in _RESOURCE_UNIT_TO_MIB.items():
            if value.endswith(suffix):
                return float(value.removesuffix(suffix)) * factor

    raise ValueError(f"Cannot parse {value_type} resource value: {value!r}")

(naming yolo'd - i am not tied to any of the names here and didn't think too hard about them)

[matt-ullmer]

Good idea — went with something close to your sketch. Pushed 2a1b013: renamed to _normalize_resource_value(value, resource_type) with a _SUFFIX_RESOURCE_TYPES = {"mem", "disk"} guard. Numeric values pass through for any type, suffix parsing only fires for mem/disk, and strings for cpu now raise ValueError. Added a test for the cpu-string case too.

[nemacysts]

do we still want to remove the unclamped_mib naming since not all of these will have an mib unit value?