2 - Avoid recreating full table mapping for every render

[cmgoffena13]

File: sqlmesh/core/snapshot/definition.py
Functions: to_table_mapping
Lines: 1994 - 2003

Issue with to_table_mapping

Issue is really in sqlmesh/core/renderer.py in _resolve_tables that we touched in Issue 1. The function calls to_table_mapping for every expression, doing a comprehension every single time. We need to cache the mapping so renders just grab the cache instead of recreating it.

Example AI Code

_TO_TABLE_MAPPING_CACHE: dict[tuple, dict[str, str]] = {}
def to_table_mapping(
    snapshots: t.Iterable[Snapshot], deployability_index: t.Optional[DeployabilityIndex]
) -> t.Dict[str, str]:
    snapshots = list(snapshots)
    deployability_index = deployability_index or DeployabilityIndex.all_deployable()
    key = (
        deployability_index.indexed_ids,
        deployability_index.is_opposite_index,
        deployability_index.representative_shared_version_ids,
        tuple((s.snapshot_id, s.version, s.dev_version) for s in snapshots),
    )
    cached = _TO_TABLE_MAPPING_CACHE.get(key)
    if cached is None:
        cached = {
            snapshot.name: snapshot.table_name(deployability_index.is_deployable(snapshot))
            for snapshot in snapshots
            if snapshot.version and not snapshot.is_embedded and snapshot.is_model
        }
        _TO_TABLE_MAPPING_CACHE[key] = cached
    return dict(cached)

Note: lru_cache on the function directly does not work — snapshots is an unhashable iterable of objects.

[cmgoffena13]

Cleaner approach is caching this all the way up inside SnapshotEvaluator.promote -- sqlmesh/core/snapshot/evaluator.py and pass into _promote_snapshot so the cache is not created per thread. Either is acceptable right now, just if you want to increase scope.