Subclass IOManager and implement handle_output(context, obj) to write the asset's output and load_input(context) to read it back
Use context.asset_key, context.partition_key, and context.metadata to construct a deterministic storage path such as s3://bucket/asset_name/partition_key/data.parquet
Register the IO manager as a resource in the Definitions object under 'io_manager' or a named key, then reference it per-asset with io_manager_key='my_io_manager'
Return the loaded object from load_input() as the expected Python type; use context.dagster_type to validate type compatibility at runtime
Add retry logic inside handle_output for transient storage errors; Dagster will not automatically retry IO manager calls
Test the IO manager with build_input_context and build_output_context helpers to verify read/write behavior without running a full Dagster pipeline
Known gotchas
If an asset has no return type annotation and the IO manager's load_input returns a non-trivial object, Dagster will still call load_input even if no downstream asset uses the output — guard against unnecessary reads
Partitioned assets require the IO manager to handle context.has_partition_key being False during non-partitioned test runs; always check before using context.partition_key
The default IO manager pickles outputs to local disk; forgetting to override io_manager_key means large DataFrames silently pickle rather than use your custom backend
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp