mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-08-21 07:02:25 +02:00
9abee46990
## Summary Builds on #7586. Adds `snapshot_frequency: int | None` to `DeltaChannel`, letting users trade storage for bounded read depth. Also promotes `channels/_delta.py` from private to public (`channels/delta.py`). ### How it works Every Nth **pregel step**, `create_checkpoint` writes a `_DeltaSnapshot` blob instead of `DELTA_SENTINEL`. The ancestor walk in `_get_channel_writes_history` terminates at the snapshot rather than walking the full chain, bounding replay to at most N steps. Snapshots are **eager**: fired even on steps where the channel had no write (via a `get_next_version` version bump), so the depth bound holds unconditionally — no risk of the cadence drifting if a channel happens to be silent at a snapshot step. ### Storage formula | Mode | Blob storage | Read depth | |------|-------------|------------| | `snapshot_frequency=None` (pure delta) | O(N) — sentinels only | O(N) steps | | `snapshot_frequency=K` | O(N²/K) — periodic snapshots of growing size | O(K) steps | | add_messages / BinOp | O(N²) — full blob every step | O(1) | At N turns with ~400 char/msg messages, total snapshot storage ≈ N²/(2K) × avg_msg_size, since each snapshot blob grows linearly with accumulated messages. ### Key design decisions - **Step-based**: `snapshot_frequency=K` means "snapshot every K pregel steps." `create_checkpoint` has the step number; the channel itself doesn't need to track writes. - **Eager**: version-bumped via `get_next_version` even on non-write steps so `put()` always stores the blob. - **`_DeltaSnapshot` NamedTuple + msgpack ext type** (`EXT_DELTA_SNAPSHOT = 7`): serde type tag dispatches in `from_checkpoint` — no dict key inspection, no collision risk. - **`from_checkpoint` semantics**: `_DeltaSnapshot` → restore value directly (no replay needed); `DELTA_SENTINEL` / `MISSING` → replay from ancestor writes; plain value → pre-migration BinOp blob. - **InMemorySaver and PostgresSaver updated**: `_get_channel_writes_history` collects the snapshot ancestor's pending_writes before terminating (they encode the *next* step's transition, unlike pre-delta migration blobs which subsume their own writes). - **`snapshot_frequency=None`** is the pure-delta default (replaces `math.inf`). ### Benchmark results (InMemory, ~400 char/msg) **Storage** | turns | ctx | freq=1 | freq=5 | freq=10 | freq=50 | freq=inf | |------:|----:|-------:|-------:|--------:|--------:|---------:| | 50 | ~10K tok | 5.9 MB | 1.2 MB | 601.3 KB | 119.8 KB | 29.5 KB | | 100 | ~20K tok | 23.7 MB | 4.8 MB | 2.4 MB | 475.8 KB | 58.4 KB | | 200 | ~40K tok | 94.6 MB | 19.0 MB | 9.5 MB | 1.9 MB | 116.4 KB | | 500 | ~100K tok | 591.5 MB | 118.4 MB | 59.2 MB | 11.8 MB | 290.3 KB | **Read latency** (avg of 5 `get_state` calls) | turns | ctx | freq=1 | freq=5 | freq=10 | freq=50 | freq=inf | |------:|----:|-------:|-------:|--------:|--------:|---------:| | 50 | ~10K tok | 0.4ms | 0.4ms | 0.7ms | 0.9ms | 1.8ms | | 100 | ~20K tok | 0.7ms | 0.9ms | 1.0ms | 1.7ms | 5.7ms | | 200 | ~40K tok | 1.5ms | 1.7ms | 4.5ms | 3.7ms | 20.1ms | | 500 | ~100K tok | 3.6ms | 4.2ms | 4.4ms | 9.0ms | 110.3ms | **Per-invoke write latency** | turns | ctx | freq=1 | freq=5 | freq=10 | freq=50 | freq=inf | |------:|----:|-------:|-------:|--------:|--------:|---------:| | 50 | ~10K tok | 1.5ms | 1.1ms | 1.1ms | 1.3ms | 1.7ms | | 100 | ~20K tok | 2.5ms | 1.6ms | 1.5ms | 1.7ms | 3.3ms | | 200 | ~40K tok | 3.4ms | 2.3ms | 2.2ms | 2.5ms | 8.3ms | | 500 | ~100K tok | 6.2ms | 4.2ms | 3.6ms | 4.1ms | 39.2ms | ## Test plan - [x] `make format` / `make lint` clean across `langgraph`, `checkpoint`, `checkpoint-postgres` - [x] `tests/test_channels.py` — 37 passing including step-based and eager-snapshot tests - [x] `tests/test_delta_channel_migration.py` — all passing - [x] Full suite: 1387 passing, 6 pre-existing failures unrelated to this branch --------- Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>