Flow Recovery Agentic AI (Quarkus Flow + LangChain4j + Persistence)

This project demonstrates workflow persistence and recovery using Quarkus Flow with Agentic LangChain4j across multiple persistence backends:

• File (MVStore)
• Redis
• PostgreSQL (JPA)

The objective is to validate that a workflow can resume execution from the last persisted step after a crash, without re-executing completed steps (saving high cost of calling again to the model).

---

🚀 Overview

The workflow is composed of three steps:

1. RiskClassifierAgent

   • Invokes an LLM (via LangChain4j + Ollama)
   • Produces a risk classification (LOW, HIGH, …)
   • Persists the result

2. CrashOnceService

   • Forces a JVM crash only once
   • Simulates a failure after persistence

3. FormatterAgent

   • Uses previously persisted data
   • Produces the final result (COMPLETED)

---

🧠 Recovery Scenario

Expected execution:

1. Workflow starts
2. Step 1 completes and is persisted
3. JVM crashes before Step 3
4. Application restarts
5. Workflow state from first agent is restored from persistence
6. Execution resumes at Step 3
7. Workflow completes successfully

---

🔄 Sequence & Recovery Flow

---

⚙️ Requirements

• Java 21+
• Maven
• Docker
• Ollama running locally

---

🐳 Dependencies

Start required services:

docker compose up -d

Run Ollama:

ollama serve
ollama pull llama3.2:3b

---

▶️ Run the application

Build

Build depending on the persistence type:

mvn clean package -Pfile -Dquarkus.profile=file
mvn clean package -Predis -Dquarkus.profile=redis
mvn clean package -Pjpa -Dquarkus.profile=jpa

Reset crash marker

In case of previous executions, be sure there's no crash marker first time:

rm -f target/agentic-crash-once.marker

Run

Run with desired persistence backend:

Redis

java -Dquarkus.profile=redis -jar target/quarkus-app/quarkus-run.jar

PostgreSQL (JPA)

java -Dquarkus.profile=jpa -jar target/quarkus-app/quarkus-run.jar

File (MVStore)

java -Dquarkus.profile=file -jar target/quarkus-app/quarkus-run.jar

---

🧪 Trigger the workflow

curl -X POST http://localhost:8080/workflow/order-agentic \
  -H 'Content-Type: application/json' \
  -d '{"orderId":"crash-1","amount":40,"customerId":"cust-1"}'

First execution:

curl: (52) Empty reply from server

This is expected — the JVM crashes intentionally.

---

🔁 Verify recovery

Restart the application:

java -Dquarkus.profile=<profile> -jar target/quarkus-app/quarkus-run.jar

Expected logs after restart (there's no "[STEP 1 - RiskAgent] output=..."):

[STEP 2 - CrashOnce] recoveredInput=...
[STEP 3 - FormatterAgent] recoveredInput=...
[STEP 3 - FormatterAgent] output={..., status=COMPLETED}

---

🗄️ Persistence backends

Redis

• Key-based storage
• Fast recovery
• Example:

<instanceId>:do/0/riskAgent

---

PostgreSQL (JPA)

• Relational storage

• Tables:

  • ProcessInstanceEntity
  • TaskInfoEntity

⚠️ Important:

quarkus.hibernate-orm.database.generation=update

---

File (MVStore)

• Embedded persistence
• Stored locally on disk
• No external dependencies required
• Useful for lightweight recovery testing

---

🔍 Debugging persistence

Redis

redis-cli --scan --pattern '*:do/*'

---

PostgreSQL

docker exec -it flow-postgres psql -U flow -d flow

SELECT * FROM ProcessInstanceEntity;

---

File (MVStore)

Inspect the data directory:

ls target/

Look for MVStore files storing workflow state.

---

🧪 What this project validates

• Step-level persistence
• Crash recovery correctness
• No re-execution of completed steps
• Agentic workflow compatibility with persistence
• Consistency across different persistence backends

---

⚠️ Known behaviors

• First request crashes intentionally
• HTTP response is interrupted
• Recovery happens automatically on restart

---

📌 Summary

This project demonstrates that:

Agentic workflows can be made fault-tolerant and resumable using Quarkus Flow persistence across Redis, JPA, and file-based storage.

---