## Summary
- Promotes the private K-channel batched ancestor-walk to a stable
public `get_delta_channel_history` / `aget_delta_channel_history` API on
`BaseCheckpointSaver` (returns `Mapping[str, DeltaChannelHistory]`, a
TypedDict with `writes` always present and `seed` `NotRequired`)
- Removes `DELTA_SENTINEL` / `_DeltaSentinel` entirely — the saver layer
is now delta-agnostic on both write and read paths
- Reworks `DeltaChannel` snapshot cadence from "every Nth superstep" to
"every N updates to this channel," persisted in
`CheckpointMetadata.delta_updates_since_snapshot`
- Adds Postgres optimizations: paged stage-1 with cursor (1024-row
pages) and per-channel UNION ALL stage-2 (no over-fetch when channels
have different chain depths)
- Default `snapshot_frequency` becomes a positive int (default `1000`);
the previous `None` opt-out is removed
## Public API
```python
class DeltaChannelHistory(TypedDict):
writes: list[PendingWrite] # always present, possibly empty
seed: NotRequired[Any] # absent if walk reached root
def get_delta_channel_history(
self, *, config: RunnableConfig, channels: Sequence[str]
) -> Mapping[str, DeltaChannelHistory]: ...
async def aget_delta_channel_history(
self, *, config: RunnableConfig, channels: Sequence[str]
) -> Mapping[str, DeltaChannelHistory]: ...
```
`config` and `channels` are keyword-only so later additions (e.g.
`page_size`) don't shift the positional API.
The TypedDict-with-`NotRequired[seed]` shape matches the existing
checkpoint-package convention (`CheckpointMetadata` is
`TypedDict(total=False)`) — absence-via-key-omission rather than
introducing a new sentinel. Pregel translates `"seed" not in hist` to
`MISSING` on its side at consume time.
The default impl walks `get_tuple` + `parent_config` correctly but is
slow on long chains; savers that care override (`InMemorySaver`,
`PostgresSaver`).
## Sentinel removal
`DELTA_SENTINEL` and `_DeltaSentinel` are deleted entirely. The saver
layer becomes delta-agnostic:
- `DeltaChannel.checkpoint()` returns `MISSING` for non-snapshot steps;
pregel's `create_checkpoint` skips MISSING so delta channels without a
snapshot simply don't appear in `channel_values`
- `InMemorySaver.put` and Postgres `put` no longer filter sentinels
(they have nothing to filter)
- `_needs_replay` becomes `stored is MISSING`
- `DeltaChannel.from_checkpoint` accepts: `MISSING` → empty,
`_DeltaSnapshot(value)` → snapshot value, plain value → pre-migration
legacy
## Snapshot cadence
`DeltaChannel.snapshot_frequency: int` (default `1000`, positive). The
previous `None` opt-out is gone.
```python
def should_snapshot(ch_name, ch):
if force_delta_snapshot: # durability="exit"
return True
return updates_since_snapshot.get(ch_name, 0) >= ch.snapshot_frequency
```
Per-channel update counters are persisted in
`CheckpointMetadata.delta_updates_since_snapshot` (`NotRequired`,
`total=False`). The counter is incremented by `_put_checkpoint` for any
delta channel in `updated_channels` and reset to `0` by
`create_checkpoint` for channels that fire a snapshot this step.
Version-format-independent — works for `int`, `float`, and `str`
versioning schemes alike.
## Postgres optimization
Two improvements internal to the override:
**Stage-1 paged with cursor** (`LIMIT 1024` internal const, `AND
checkpoint_id < ?` for subsequent pages). The previous unpaged form
scanned every checkpoint in `(thread_id, ns)` and was pathological at
high thread depths.
**Stage-2 per-channel UNION ALL**: one `WHERE channel='X' AND
checkpoint_id = ANY(chain_X)` branch per channel plus one seed-blob
branch per channel with a seed. The previous form filtered by `channel =
ANY(channels) AND checkpoint_id = ANY(union_chain_cids)`, over-fetching
writes when channels had different chain depths (`K ×
max(chain_lengths)` vs the correct `sum(chain_lengths)`).
Both improvements stay internal to `PostgresSaver`/`AsyncPostgresSaver`;
the public contract returns a single `Mapping`.
## Benchmarks
`libs/langgraph/tests/test_delta_channel_benchmark.py`. Run via `python
libs/langgraph/tests/test_delta_channel_benchmark.py`. Postgres against
local pg:5441.
Results below trimmed to the high-signal cells. Sub-millisecond /
sub-100-turn rows omitted as warmup-bound; freq=1 omitted (chain depth =
1, nothing to optimize); peak read-time memory and Postgres storage are
flat between branches and omitted. Deep-thread reads and the
cadence-rework storage win are the load-bearing numbers.
### Postgres reads, 500 turns
| Scenario | main | branch | Δ |
|---|---:|---:|---:|
| Single-channel deep read | 17.7 ms | **6.1 ms** | **-66%** |
| Single-channel, 1000 turns | 35.0 ms | **14.3 ms** | **-59%** |
| K=3 channels, freq=50 uniform | 70.5 ms | **41.4 ms** | **-41%** |
| K=8 channels, freq=50 uniform | 214.2 ms | **139.4 ms** | **-35%** |
| K=8 channels, mixed freq (25/50/100/.../1000) | 295.6 ms | **214.4
ms** | **-27%** |
K-channel batching + paged stage-1 + per-channel UNION ALL stage-2 doing
exactly what they should at depth.
### InMemory reads, 500 turns
| Scenario | main | branch | Δ |
|---|---:|---:|---:|
| Single-channel deep read | 7.9 ms | **3.8 ms** | **-52%** |
| Single-channel, 1000 turns | 15.6 ms | **7.2 ms** | **-54%** |
| K=8 channels, freq=50 uniform | 112.3 ms | 94.6 ms | -16% |
| K=8 channels, mixed freq | 184.9 ms | **134.5 ms** | **-27%** |
### InMemory storage, 500 turns (cadence-rework win)
| Scenario | main | branch | Δ |
|---|---:|---:|---:|
| K=3, freq=50 uniform | 8.7 MB | **3.3 MB** | **-62%** |
| K=3 mixed freq | 3.8 MB | **1.3 MB** | **-66%** |
| K=8, freq=50 uniform | 23.1 MB | **8.7 MB** | **-62%** |
| K=8 mixed freq | 11.5 MB | **4.2 MB** | **-64%** |
Snapshot frequency now counts **channel updates** instead of
**supersteps**. On graphs where supersteps outpace per-channel updates
(e.g., input/end steps that don't write to channels), branch stores ~3×
fewer snapshot blobs.
### Tradeoff worth flagging
InMemory K=3 with mixed frequencies (50/200/1000) at 500 turns: **+64%
read latency** (46.6 → 76.5 ms). The mixed scenario has a channel with
`freq=1000` that goes the entire 500-turn run with no snapshot. On main,
the old superstep-counted cadence happened to fire at step=500 anyway.
New cadence gives users explicit control over walk depth via
`snapshot_frequency`. The K=8 mixed case still wins overall (-27%); this
regression is specific to the K=3 mixed shape.
Default `snapshot_frequency=1000` is the upper bound on walk depth —
it's a tunable knob.
## Tests
- New sqlite smoke test (`test_get_delta_channel_history.py`) exercises
the inherited default `BaseCheckpointSaver` impl via `SqliteSaver` /
`AsyncSqliteSaver` end-to-end with a real `DeltaChannel`-backed graph.
Sqlite uses the default unchanged — this validates the default path
actually works on a real second saver, not just on the optimized
override.
- Module-level `pytest.importorskip("langgraph.channels.delta")` guards
the test for sqlite's standalone CI environment (matches the postgres
pattern).
## Test plan
- [x] `libs/checkpoint`: 150 passed, 16 skipped
- [x] `libs/langgraph` (channels + delta migration): 41/41 (post-merge)
- [x] `libs/langgraph` (full pregel suite): 1784 passing — 6 "failures"
verified via `env -i` clean shell are local LangSmith env vars + `git
describe revision_id` polluting LangChain metadata fixtures; CI is
unaffected
- [x] `libs/checkpoint-postgres`: 40/40 saver tests + 3/3 delta channel
reconstruction tests against local Postgres
- [x] `libs/checkpoint-sqlite`: 105/105 (incl. retry-passed flake
`test_ttl_refresh`, unrelated to this PR)
- [x] Lint clean across all four libs (`ruff format`, `ruff check`,
`mypy`)
- [x] Branch-vs-main benchmarks — see results above
---------
Co-authored-by: Quanzheng Long <long@langchain.dev>
Co-authored-by: Cursor <cursoragent@cursor.com>
## Summary
Replaces the single-roundtrip `UNION ALL` DeltaChannel read with a
two-stage query that avoids fetching unused snapshot blobs, then removes
the old combined path entirely.
### Problem
`_get_channel_writes_history` used a single `UNION ALL` query that
fetched **all** checkpoint metadata, writes, and blobs for a
`(thread_id, channel)` in one shot. With `snapshot_frequency=N`, this
pulled back O(N/freq) full-size snapshot blobs even though only the
nearest one is needed to seed reconstruction. At 500 turns with
`snapshot_frequency=10`, this meant fetching ~100 complete
message-history snapshots per read.
### Solution
Two-stage read:
- **Stage 1** — lightweight scan of `checkpoints` only (no blob bytes):
walks the parent chain from the target checkpoint and stops at the first
ancestor with a snapshot, returning `chain_cids` and `seed_version`
- **Stage 2** — targeted fetch: only the writes for `chain_cids` and the
single seed blob at `seed_version`
The two-stage path is now unconditional — the old combined query and
`LG_DELTA_TWO_STAGE_QUERY` env-var gate have been removed.
### Sentinel cleanup
`DELTA_SENTINEL` is now a pure in-memory signal and is never written to
storage:
- Postgres `put()` already stripped it from `channel_values` before
writing blobs
- Memory saver `put()` now stores `"empty"` instead of serializing the
sentinel
- `EXT_DELTA_SENTINEL` (msgpack ext code 8) removed from
`JsonPlusSerializer`
- `DELTA_SENTINEL` is kept as an in-memory marker:
`DeltaChannel.checkpoint()` returns it so savers know to skip it, and
`_ChannelWritesHistory.seed` uses it to mean "no snapshot found, start
from empty"
## Performance
Benchmarked at `snapshot_frequency=10` on Postgres (`~100 tok/msg`):
| turns | old combined query | two-stage |
|------:|-------------------:|----------:|
| 50 | 6.0ms | 2.8ms (2.1x faster) |
| 100 | 10.1ms | 5.6ms (1.8x faster) |
| 500 | **216.1ms** | 15.3ms (**14x faster**) |
The old query's read time grew super-linearly with turn count because
each read fetched O(N/freq) full snapshot blobs. Two-stage keeps read
depth bounded by `snapshot_frequency` regardless of thread length.
## Test plan
- `make test` in `libs/checkpoint`, `libs/checkpoint-postgres`,
`libs/langgraph`
- Removed `test_delta_sentinel_serde_round_trip` (sentinel no longer
serializable)
- Updated `test_memory.py` — delta channel blobs stored as `"empty"`,
not serialized sentinel
- Updated `test_channels.py` — `channel_values` no longer contains
sentinel key for DeltaChannels
- Deleted `test_delta_channel_two_stage_benchmark.py` (one-stage vs
two-stage comparison; path no longer exists)
---------
Co-authored-by: Sydney Runkle <54324534+sydney-runkle@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- Leave it up to each checkpointer implementation to decide whether to merge in configurable/metadata (previously PregelLoop would do some of this always)
- Never copy over internal langgraph keys into checkpoint.metadata (these are redundant/misleading to include)
Prepare langgraph-checkpoint for 0.5
- Given we have no upper bound on langgraph-checkpoint dep need to undo all changes in langgraph-checkpoint that might break previous versions of langgraph
- This has been superseded by saving the individual writes of each task through put_writes()
- Removing this speeds up checkpoint operations as it was duplicating data saved elsewhere already
- Instead store sends in a Topic channel, removing the need to fetch sends as writes against the parent checkpoint
- Remove deprecated/unused functions in langgraph-checkpoint (will require bumping min range for langgraph-checkpoint in langgraph lib)
- Implement migration of old pending sends in langgraph-checkpoint-postgres
- Ensure parent config of `checkpoint_during=False` checkpoints always points to checkpoints that were also saved
- This makes our checkpoint benchmarks more closely resemble the behavior of our prod checkpointers
- Also found and fixed a bug w multiple subgraphs in same node accidentally sharing checkpoints
(Keep MemorySaver around for backwards compatibility)
"MemorySaver" is ambiguous: is it saving memories? Where is it saving
memories to?
InMemorySaver aligns naming InMemoryStore as well as similar LangChain
objects (InMemoryVectorStore, etc.)
