Handling Overwrites in Span Metadata: Best Practices

Deepankar M. · 2025-10-13T12:46:32.260Z

Hi, we use span.set_attribute(SpanAttributes.METADATA, json.dumps(...)) to set metadata. Just found that any subsequent call to span.set_attribute(SpanAttributes.METADATA, json.dumps(...)) will override the previous metadata. We cannot combine these as they are from 2 unrelated parts of the code operating on the same span. How should we handle this?

Handling Overwrites in Span Metadata: Best Practices | Arize AI Community

6 comments

· Sorted by Oldest

RunLLM
APP
·
·
Setting `span.set_attribute(SpanAttributes.METADATA, ...)` multiple times will overwrite the previous value, as span attributes are key-value pairs and not merged automatically. There is no built-in support in Phoenix or OpenInference for merging metadata from multiple sources into a single attribute; each call replaces the previous value (<https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>, <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md|OpenInference context managers>). To avoid overwriting, you must aggregate all metadata into a single dictionary before setting the attribute, or use distinct keys for each metadata source (e.g., `metadata.part1`, `metadata.part2`). There is no documented mechanism to merge or append to the same attribute key from different code locations (<https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>). Would you like more detail or example code for aggregating metadata? Sources: - <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3> - <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md|https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
RunLLM attributes are not overwritten with subsequent calls right? Only metadata is overwritten?
RunLLM
APP
·
·
All span attributes—including metadata—are overwritten on subsequent calls to `span.set_attribute(key, value)` for the same key. This is standard OpenTelemetry and Phoenix behavior: setting the same attribute key again replaces the previous value, regardless of whether the key is `metadata` or any other attribute (<https://arize-ai.github.io/phoenix/interfaces/_arizeai_phoenix-client.__generated___api_v1.components.html#description-73|OpenTelemetry attribute spec>, <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>). So, any attribute (not just metadata) will be overwritten if set multiple times with the same key. Would you like more detail or examples? Sources: - <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3> - <https://arize-ai.github.io/phoenix/interfaces/_arizeai_phoenix-client.__generated___api_v1.components.html#description-73|OpenTelemetry attribute spec>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
I see, so this gets overwritten since SpanAttribute.METADATA is used as the key. Makes sense. RunLLM (or any humans 🙇) are there examples of how to query in phoenix using custom attributes? Because of the overwriting issue, I cannot use metadata for filtering and I can't find any example in phoenix docs to filter using custom attributes.
RunLLM
APP
·
·
Yes, you can filter spans in Phoenix using any custom attribute by referencing its key directly in the query/filter expression. For example, if you set a custom attribute like `span.set_attribute("customerID.name", "12345")`, you can filter with `customerID.name == "12345"` in the Phoenix UI or in a `SpanQuery` (<https://arize.com/docs/phoenix/tracing/how-to-tracing/custom-spans#add-attributes-to-a-span|custom spans docs>, <https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui|filter spans docs>). Example: ``` from phoenix.trace.dsl import SpanQuery query = SpanQuery().where('customerID.name == "12345"') df = px.Client().query_spans(query) ``` Or in the UI search bar: `customerID.name == "12345"` Would you like more detail or more code examples? Sources: - <https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui|https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
ah nice! This seems to work. Thank you 🤖
🙌1

RunLLM
APP
·
·
Setting `span.set_attribute(SpanAttributes.METADATA, ...)` multiple times will overwrite the previous value, as span attributes are key-value pairs and not merged automatically. There is no built-in support in Phoenix or OpenInference for merging metadata from multiple sources into a single attribute; each call replaces the previous value (<https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>, <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md|OpenInference context managers>). To avoid overwriting, you must aggregate all metadata into a single dictionary before setting the attribute, or use distinct keys for each metadata source (e.g., `metadata.part1`, `metadata.part2`). There is no documented mechanism to merge or append to the same attribute key from different code locations (<https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>). Would you like more detail or example code for aggregating metadata? Sources: - <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3> - <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md|https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/how-to-tracing/add-metadata/customize-spans.md>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
RunLLM attributes are not overwritten with subsequent calls right? Only metadata is overwritten?
RunLLM
APP
·
·
All span attributes—including metadata—are overwritten on subsequent calls to `span.set_attribute(key, value)` for the same key. This is standard OpenTelemetry and Phoenix behavior: setting the same attribute key again replaces the previous value, regardless of whether the key is `metadata` or any other attribute (<https://arize-ai.github.io/phoenix/interfaces/_arizeai_phoenix-client.__generated___api_v1.components.html#description-73|OpenTelemetry attribute spec>, <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|customize spans docs>). So, any attribute (not just metadata) will be overwritten if set multiple times with the same key. Would you like more detail or examples? Sources: - <https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3|https://arize.com/docs/phoenix/tracing/how-to-tracing/add-metadata/customize-spans#python-3> - <https://arize-ai.github.io/phoenix/interfaces/_arizeai_phoenix-client.__generated___api_v1.components.html#description-73|OpenTelemetry attribute spec>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
I see, so this gets overwritten since SpanAttribute.METADATA is used as the key. Makes sense. RunLLM (or any humans 🙇) are there examples of how to query in phoenix using custom attributes? Because of the overwriting issue, I cannot use metadata for filtering and I can't find any example in phoenix docs to filter using custom attributes.
RunLLM
APP
·
·
Yes, you can filter spans in Phoenix using any custom attribute by referencing its key directly in the query/filter expression. For example, if you set a custom attribute like `span.set_attribute("customerID.name", "12345")`, you can filter with `customerID.name == "12345"` in the Phoenix UI or in a `SpanQuery` (<https://arize.com/docs/phoenix/tracing/how-to-tracing/custom-spans#add-attributes-to-a-span|custom spans docs>, <https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui|filter spans docs>). Example: ``` from phoenix.trace.dsl import SpanQuery query = SpanQuery().where('customerID.name == "12345"') df = px.Client().query_spans(query) ``` Or in the UI search bar: `customerID.name == "12345"` Would you like more detail or more code examples? Sources: - <https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui|https://arize.com/docs/phoenix/how-to/extract-data-from-spans#how-to-apply-filters-ui>
💡 Hint: Mention RunLLM in the thread for followups.
Deepankar M.
·
·
ah nice! This seems to work. Thank you 🤖
🙌1