It seems likely that Phoenix does not perform automatic deduplication of identical input/output data across deeply-nested spans; each span's attributes (including large context or results) are stored as-is, even if repeated in the trace. There is no mention of deduplication logic for span input/output in the available documentation or code discussions for Mastra or Phoenix in general. If you have repeated large payloads, they will be stored for each span that includes them, potentially increasing storage usage. Would you like more detail on how to manage or mitigate this?
Regarding Postgres, my understanding is that Phoenix relies on Postgres' built-in TOAST (The Oversized-Attribute Storage Technique) for handling large row data. When a row exceeds 8KB, Postgres automatically TOASTs the data, storing large values out-of-line and keeping a pointer in the main table, so you do not need to configure this separately for Phoenix. This is standard Postgres behavior and is not altered by Phoenix's persistence layer configuration. Would you like more technical detail on how TOAST works or how Phoenix interacts with it?
Sources:
- <https://github.com/Arize-ai/phoenix/pull/5250|https://github.com/Arize-ai/phoenix/pull/5250>
