Advanced10 min

RetryPolicy, Error Taxonomy & CachePolicy

Structured error handling in LangGraph: RetryPolicy for transient failures, error taxonomy for routing errors to the right handler, and CachePolicy for avoiding redundant computation.

Quick Reference

→RetryPolicy on nodes: max_attempts, retry_on, initial_interval, backoff_factor for transient errors
→Error taxonomy: transient → RetryPolicy, LLM-recoverable → loop back, user-fixable → interrupt(), unexpected → bubble up
→CachePolicy on @task: TTL-based caching to avoid recomputing identical operations
→InMemoryCache or custom cache backends for persistent caching across runs
→RetryPolicy works on both graph nodes (add_node) and @task decorators
→Combine all three for production-grade error handling

RetryPolicy on Nodes

Adding retry policies to graph nodes

Parameter	Type	Default	Purpose
max_attempts	int	3	Total attempts including the first try
initial_interval	float	1.0	Seconds to wait before first retry
backoff_factor	float	2.0	Multiply interval by this each retry
retry_on	tuple[Exception]	All exceptions	Only retry on these exception types
jitter	bool	True	Add random jitter to prevent thundering herd

Error Taxonomy: Route Errors to the Right Handler

Classify errors by who can fix them → route to the right handler

LLM-Recoverable Errors

When the LLM produces malformed output (bad JSON, missing fields, wrong format), don't retry blindly — that wastes money on the same prompt. Instead, store the error in state and loop back to the LLM with the error message. The LLM can usually fix its own mistakes when told what went wrong.

Sign in to read this article

This is a premium article. Sign in with your Google account to continue.