View a markdown version of this page

Function error codes and other causes of missing metadata - Elemental Inference

Function error codes and other causes of missing metadata

A run that fails writes nothing, so every runtime problem looks the same from GetMetadata: no item for that window, an empty items array, ResourceNotFoundException, or MetadataUnavailableException while no later run has succeeded. The Preview metadata section on the console's Feed outputs tab shows nothing for the same windows.

To find the cause, enable vended logs for the feed. A run that fails for one of the following reasons logs an ERROR record whose detail.error_code names the cause, and whose detail includes the function_id. Repeated failures with the same code are logged at most once every 5 seconds for each feed. For information about how to enable the logs and read a record, see Monitoring AWS Elemental Inference with vended logs.

error_code Cause and what to do
FUNCTION_ROLE_ACCESS_DENIED Elemental Inference couldn't assume the feed's access role. Check the role's trust policy. For more information, see Trust policy.
FUNCTION_MODEL_ACCESS_DENIED The model denied the invocation. Check that the access role allows bedrock:InvokeModel on the model, that you have access to the model in the feed's Region, and that no service control policy in your organization denies the call. An explicit deny in a service control policy can't be overridden by the role's permissions. For more information, see Permissions policy.
FUNCTION_MODEL_THROTTLED Your account's Amazon Bedrock quota for the model is exhausted. Lengthen the trigger interval, request a higher Amazon Bedrock quota, or use provisioned throughput. For more information, see Costs and throughput.
FUNCTION_MODEL_REQUEST_INVALID The model rejected the request body that the request expression built. Make the body match the request format of the model you chose. The log record doesn't include the model's message, because it can contain request content. A model that requires a data retention mode higher than your account's mode in the feed's Region also fails this way. For more information, see Data retention in the Amazon Bedrock User Guide.
FUNCTION_MODEL_TIMEOUT The model didn't respond within the invocation timeout, about 18 seconds. Simplify the request, reduce the frame count, choose a faster model, or lengthen the trigger interval.
FUNCTION_MODEL_ERROR The model returned another error. detail.status_code holds the HTTP status. Check the model's availability in your account, and its service health.
FUNCTION_REQUEST_EXPRESSION_INVALID The request expression failed on this window's data. Check that it handles missing or empty context, such as a window with no speech. For more information, see Expression limits and guardrails.
FUNCTION_RESPONSE_EXPRESSION_INVALID The response expression failed on the model's response. Temporarily set the response expression to $ to see the whole response, then fix the expression. For more information, see Response expressions.
FUNCTION_MEDIA_TOO_LARGE This window's frames encoded larger than the model accepts in one request. Reduce the frame count or set a smaller frame size.

Some failures log nothing. The most common is a run that doesn't finish within about 20 seconds of the end of its window on a live feed. If the logs are empty, work through the symptoms in the following table.

Symptom Likely cause
No metadata from any output on the feed, including the native ones. The source isn't delivering media, or the feed isn't active. This isn't a function problem. For more information, see Lifecycle of an AWS Elemental Inference workflow.
Metadata appears intermittently. Runs are taking close to the time limit, or your Amazon Bedrock throughput quota is throttling the calls. Lengthen the interval, choose a faster model, reduce the frame count, or use provisioned throughput.
Metadata appears, but the value is a whole model response. The function has no response expression, or the expression evaluated to its entire input.
Metadata appears, but a field you expected is missing. The response expression's path doesn't match the model's response shape. Temporarily set the response expression to $ to see the whole response, then fix the path. For more information, see Response expressions.
contextTypes is empty, or missing a type you expected. The request expression doesn't actually reference that media path, so Elemental Inference supplies nothing for it. Fix the expression and update the function; contextTypes is recomputed on every update. While a feed uses the function, an update can't change contextTypes.
The frame settings don't appear on the console attachment form. The form shows only the settings for the context that the function's contextTypes reports. An empty list means the expression references no media.
Metadata stops at the same moment for every function on the feed. Check whether the function outputs were disabled, or dropped by a feed update that resent an incomplete output list.
Tip

To decide whether a problem is in the model call or in the feed, attach a passthrough function with the same trigger, such as one whose request expression is { "ok": true }. A passthrough function needs no access role and invokes no model. If the feed is at its function-output quota, disable the original output first. If the passthrough function produces metadata, the feed and the trigger are working, and the problem is in the invocation or the response expression.

For a run that succeeds but returns an unexpected result, the item's bedrockRequestId identifies the model invocation. Include it when you contact AWS Support about the model's response. For more information, see Reading function results.

Function runs don't emit Amazon CloudWatch metrics to your account. For information about metrics that show whether a feed is processing at all, see Monitoring AWS Elemental Inference with Amazon CloudWatch. For information about the events that a feed emits, see Monitoring AWS Elemental Inference events in Amazon EventBridge.