Limits and quotas
Every rate limit, size limit, expiration, and constraint in Avra Core in one place, with what happens when you reach each one.
This page gathers the limits in Avra Core: rate limits, sizes, URL lifetimes, and the constraints on batches, versions, and webhooks. Each table says what happens when you reach a limit and links to the page that explains it. Read it when you size an integration, and come back to it when a request is refused.
Some limits are exact and some are approximate. The prediction rate limit is enforced at Avra’s network edge, so treat it as approximate. The others are enforced by the API itself, and the error tells you which one you reached.
Time windows at a glance
Several of Core’s limits are lengths of time. They range from the 10 seconds a key stays blocked after a rate-limited prediction to the hour an upload URL stays valid.
When a limit refuses a request, the error names it. Wait at least 10 seconds after a 429 on a prediction, and the seconds in Retry-After after a 429 on the files endpoints. Retry a timed-out prediction with backoff, split an upload request that has too many files, wait for a batch to complete or fail before you ask for its result, and request a new URL when one expires.
Live predictions
| Limit | Value | When you reach it | Details |
|---|---|---|---|
| Requests per API key | About 100 per 10 seconds | 429 with rate_limited, and the key is blocked for 10 seconds |
Live predictions |
| Records per request | One | Send more records through a batch | Batch inference |
| Time to answer | 30 seconds | 500 with prediction timed out |
Handle errors |
Batch jobs don’t count against the prediction rate limit. For anything more than a few thousand records at once, use batch inference.
Files and uploads
| Limit | Value | When you reach it | Details |
|---|---|---|---|
| Files endpoints, per organization | 1,500 requests per minute | 429 with too many requests, and a Retry-After header with the seconds to wait |
Batch inference |
| Files endpoints, per caller | 300 requests per minute for each API key, and for each member in the dashboard | The same 429 |
Batch inference |
| Files per upload request | 1,000 | 400 with too many files in one request |
Prepare the input files |
| Upload URL lifetime | 1 hour, until the expiration in the response |
The upload is refused. Request new URLs | Batch inference |
| Checksum | A base64-encoded, 128-bit MD5 digest for every file | 400 with md5 must be a base64-encoded 128-bit digest |
Prepare the input files |
The files endpoints are /v1/api/files/*: requesting upload URLs and listing uploaded files.
Batches
| Limit | Value | When you reach it | Details |
|---|---|---|---|
| Availability | Enabled per organization | Ask your Avra team to turn it on | Batch inference |
| File formats | Parquet or CSV, one format for every file in a batch | 400 with batch input files must all have the same format |
Prepare the input files |
| Parquet layout | At most 100k rows per row group and 5 GB per file, as guidance | Very large jobs may need sizing beforehand, so reach out to your Avra team | Prepare the input files |
| Reference date | A YYYY-MM-DD date. In the dashboard, up to two years back |
400 with invalid reference date |
Run a batch from the dashboard |
| Result availability | Once the batch has completed or failed | 409 with batch result is not ready |
Download the result |
| Result URL lifetime | 15 minutes, until the expires_at in the response |
Request the result again for a new URL | Download the result |
| Cancellation | Not available through the v2 API | Contact Avra support to stop a batch | Batch statuses |
Batches run in queues with resource quotas, so a batch can wait as queued before it runs.
Versions and aliases
| Limit | Value | When you reach it | Details |
|---|---|---|---|
| Aliases per version | One | 409 with model version already has an alias. Remove the version’s alias first |
Promote and roll back |
| Versions per alias | One, per model, across your organization | 409 with model alias already in use, unless you reassign it |
Aliases |
| Alias names | Lowercase letters, digits, and hyphens | Other characters are converted or dropped | Aliases |
| Versions that serve calls | Active versions only | 400 with model version is not active |
The life of a version |
Webhooks
| Limit | Value | When you reach it | Details |
|---|---|---|---|
| Delivery | At least once | The same event can arrive more than once. Deduplicate on idempotency_key |
Retries |
| Retries | With exponential backoff, after a non-2xx response or a timeout | Once the retries run out, the delivery is marked failed and can be replayed | Retries |
| Signature timestamp | Reject deliveries outside your tolerance window; 5 minutes is a sensible default | Your receiver refuses the delivery | Verify deliveries |
| Scope | Your whole organization | Endpoints receive the organization’s events | Configure endpoints |
API keys and embeddings
| Limit | Value | Details |
|---|---|---|
| API key scope | One workspace | Live predictions |
| Rotation overlap | The old key can keep working for up to 30 minutes after a rotation | API reference |
| Embedding dimensions | 1024, shortened only to 512, 256, 128, 64, 32, or 16 | Technical guide |
Requesting a higher limit
If your production traffic needs more, contact your Avra account manager with your expected request volume and traffic pattern. For large batch jobs, reach out before you send them, so your Avra team can size the job.