Lumen × LangGraph × Kova

Lumen × LangGraph × Kova

Goal: swap a LangGraph Agent's Checkpointer from SqliteSaver to LumenCheckpointer, deploy to Kova, and observe the crash-recovery comparison.

Before / After

Before — Plain LangGraph

Single machine, no remote backup · Cross-process requires locking · Millisecond crash recovery · No cost tracking.

After — Lumen + Kova

Microsecond crash recovery (Kova WAL engine) · Native multi-process safety · Automatic Trace + Cost · Remote persistence.

Before

from langgraph.checkpoint.sqlite import SqliteSaver

graph = wf.compile(checkpointer=SqliteSaver.from_conn_string("./state.db"))

After

from lumen_ai import LumenCheckpointer, LumenTracer
from kova import KovaClient

kova = KovaClient("kova://prod-cluster")

graph = wf.compile(
    checkpointer=LumenCheckpointer(kova_client=kova),
    callbacks=[LumenTracer()],
)

Local Verification

1. Start a local Kova

   docker run -d --name kova -p 9999:9999 ghcr.io/hanmahong5-arch/kova:latest

2. Run your Agent

   python my_agent.py

3. Simulate a crash

   kill -9 $(pgrep -f my_agent)

4. Restart — resume from the breakpoint

   python my_agent.py

Deploy to Kova

# agent.yaml
name: my-research-agent
entry: python my_agent.py
replicas: 3
checkpointer: lumen
wal:
  retention: 7d

lumen deploy --target kova://prod-cluster

Crash-Recovery Comparison Experiment

Metric	SqliteSaver	Lumen + Kova
Recovery latency	8ms (single machine)	Microsecond
Multiple replicas	Requires extra locking	Native
Cross-datacenter	Manual replication	Async replication built in
LLM re-invocation	Re-invokes if SQLite write fails	Never re-invokes
Trace	None	Auto-correlated

Where the microsecond recovery comes from

The Kova engine's scheduling latency is as low as 3μs (full FIFO pipeline Criterion benchmark of 3.17μs, 315K ops/s). LumenCheckpointer hands LangGraph's checkpoint writes to the Kova WAL — recovery uses engine-level WAL replay instead of SQLite file-level recovery, so it is microsecond-scale, far faster than single-machine SQLite's millisecond scale.

Next Steps

Next steps

Migrate from LangGraphLumen Python SDKKova Concepts