Reduce storage by changing cumulative_probability and event_value from float8 to float4

[Ahmad-Wahid]

Context

The TimedBelief table (see e62ac5f519d7_create_table_for_timed_beliefs.py) defines both cumulative_probability and event_value as sa.Float(), which SQLAlchemy/Postgres maps to float8 (double precision, 8 bytes per value).

This is our largest and fastest-growing table, so column width has a direct effect on disk/storage costs and I/O.

Proposal

Change cumulative_probability and event_value from float8 (double precision) to float4 (single precision), via a new Alembic migration (ALTER COLUMN ... TYPE real).

Rough estimate: this could reduce the size of the timed_beliefs table by roughly 15-20%, since these are two of the core per-row numeric columns.

Things to check before implementing

• Precision loss: float4 has ~7 significant decimal digits vs ~15-17 for float8. Need to confirm this is acceptable for:
  • event_value (sensor readings/schedules/forecasts - magnitudes and required precision vary by sensor/unit)
  • cumulative_probability (bounded in [0, 1], where lower precision may be more acceptable)
• Migration cost/downtime: ALTER COLUMN TYPE on a large table rewrites it; needs a migration strategy (e.g. batched backfill, or accept the lock/downtime) given this is likely the largest table in most deployments.
• Any downstream code that assumes float64 precision (e.g. numpy/pandas dtype expectations when reading beliefs into a BeliefsDataFrame, or serialization/rounding logic in the API).
• Whether this should be an opt-in migration or bundled with other maintenance (partitioning, etc.).