sandialabs / LoQS / 41205197238

90%

push

github

Improve shot-checkpoint reads and writes (#169)

* Memoize content hashes during checkpoint encodes

Shot-checkpoint writes rehashed each subtree at every level, so a
numpy History write took about 90x its shot; hashing is now linear and
the #153 Part 5 MWE drops from 2.29 to 0.19 s per checkpointed shot.

- Memoize Serializable._serial_hash for one top-level encode in a
  ContextVar that holds each object and never stores a cyclic hash.
- Hash non-object numpy arrays from shape, dtype and raw bytes, so
  equal values of different dtypes no longer dedupe as copies.
- Decode HDF5 complex arrays bit-exactly through the ndarray.imag
  setter, so NaN, inf and -0.0 survive and decoded seeds match.
- Minor: skip the hash when there's no encode cache; tests for memo
  scope, id reuse, dtypes and shapes, call scaling, complex round trip.

Co-authored-by: claude-opus-5-5 <noreply@anthropic.com>
Co-authored-by: claude-sonnet-5-5 <noreply@anthropic.com>

* Skip needless decode walks in shot-checkpoint reads

Reading shots back re-walked every decoded object for placeholders and
scanned a noise model before a program's blob, so a numpy LoGST read
cost 1.7x its shots; it now costs about half (22.8 s to 6.1 s).

- Run Serializable.decode's placeholder walk only once a decode in the
  context has handed out a DeferredRef, counted in a ContextVar.
- Search a group's collapsed blob before its real children when
  resolving a missing source; cache IDs are unique in a file.
- Minor: drop a redundant DeferredRef branch in JSON decode_cached_obj;
  test a placeholder reached by a later decode.

Co-authored-by: claude-sonnet-5-5 <noreply@anthropic.com>
Co-authored-by: claude-opus-5-5 <noreply@anthropic.com>

---------

Co-authored-by: claude-opus-5-5 <noreply@anthropic.com>

3701 of 4430 branches covered (83.54%)

Branch coverage included in aggregate %.

11975 of 13165 relevant lines covered (90.96%)

0.91 hits per line