dcdpr / jp / 27723188554

67%

push

github

perf(storage, workspace): Batch-load conversation metadata (#760)

`load_conversation_metadata` rescanned the conversations directory for
every id, making the full index load quadratic in the number of
conversations. This commit introduces a batch variant that enumerates
each storage root once and resolves all requested ids in a single pass.

`load_conversation_metadata_batch` is added to the `LoadBackend` trait
with a serial default so existing backends stay correct without changes.
`FsStorageBackend` overrides it to delegate to the underlying `Storage`,
which builds a `HashMap<ConversationId, (path, is_user)>` in one
directory scan and then fans out via `par_iter`.

Ids that are not found in the scan (e.g. in-flight persists that renamed
the directory after the scan) fall back to the single-id loader,
preserving the archive retry and the in-flight-persist retry logic.

`load_conversation_metadata_at` is extracted as a shared helper used by
both the batch and the existing single-id paths, removing the duplicate
path-resolution logic.

`load_count_and_timestamp_events` is also reworked: instead of
deserializing the full event array into a `Vec`, it now drives a
streaming `EventSummary` visitor that pulls one element at a time and
discards all but the last timestamp, keeping peak memory flat regardless
of how many events a conversation has.

`Workspace::initialize_conversations` is simplified accordingly; the
manual `par_iter` + `filter_map` + secondary loop is replaced with a
single `for` loop over the batch results.

---------

Signed-off-by: Jean Mertz <git@jeanmertz.com>

45 of 102 new or added lines in 4 files covered. (44.12%)

35658 of 53540 relevant lines covered (66.6%)

407.4 hits per line