Skip to content

Resuming a pipeline from a snapshot file turns int dict keys into strings #13492

Closed

Description

@RyanLin5967

Describe the bug

Currently, _save_pipeline_snapshot saves the snapshot with json.dump, which writes int keys as JSON strings, and the saved schema records only the values' types, so the resumed component gets '1' where the run had 1. Resuming from the in-memory e.pipeline_snapshot keeps the int keys.

Expected behavior

The resume returns {'sink': {'v': {1: 'a', 2: 'b'}}}, the same as a run without the breakpoint.

Additional context

A snapshot saved when a run fails does the same. Float, bool and None keys come back as '1.5', 'true' and 'null', and {1: 'a', '1': 'b'} as {'1': 'b'}.

To Reproduce

import os
os.environ["HAYSTACK_PIPELINE_SNAPSHOT_SAVE_ENABLED"] = "true"

from haystack import Pipeline, component
from haystack.core.errors import BreakpointException
from haystack.core.pipeline.breakpoint import load_pipeline_snapshot
from haystack.dataclasses.breakpoints import Breakpoint

@component
class Src:
    @component.output_types(v=dict[int, str])
    def run(self):
        return {"v": {1: "a", 2: "b"}}

@component
class Sink:
    @component.output_types(v=dict[int, str])
    def run(self, v: dict[int, str]):
        return {"v": v}

p = Pipeline()
p.add_component("src", Src())
p.add_component("sink", Sink())
p.connect("src.v", "sink.v")

try:
    p.run({}, break_point=Breakpoint("sink", snapshot_file_path="snapshots"))
except BreakpointException as e:
    print(p.run({}, pipeline_snapshot=load_pipeline_snapshot(e.pipeline_snapshot_file_path)))
    # {'sink': {'v': {'1': 'a', '2': 'b'}}}, should be {'sink': {'v': {1: 'a', 2: 'b'}}}
    # the docs say the JSON snapshot has "the same information as in the BreakpointException"

FAQ Check

System

  • OS: macOS 26.7.1, arm64 (Darwin 25.6.0)
  • Haystack version: 3.4.1 and main

Activity

  1. added theissue type on Oct 30, 2026
  2. rupak-eng commented on Oct 30, 2026

    @rupak-eng

    Fix opened in #13497: dicts with non-string keys in snapshot payloads are now serialized as explicit key/value entries (each key keeps its own schema), so dict[int, str] values survive a snapshot-file round-trip with their key types intact. String-keyed dicts keep the existing properties format unchanged, and snapshots written before this change still load as before. Includes the issue's end-to-end repro as a regression test plus JSON round-trip unit tests.

  3. added
    P2Medium priority, add to the next sprint if no P1 available
    on Nov 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

P2Medium priority, add to the next sprint if no P1 available

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions