Remote inputs

Submit production CSV or Parquet inputs by signed HTTPS URL.

Remote inputs

Pass an HTTP(S) URL as data to avoid loading a large file into the SDK process. The URL must return the object with an unauthenticated GET; s3:// URIs are not accepted. Downloads use only the URL’s authorization; Sutro never signs caller-selected objects with service credentials. Private S3/R2 objects therefore require a caller-authorized presigned GET URL. Public objects may use an unsigned URL.

Destinations must resolve only to public IPv4/IPv6 addresses. HTTP uses port 80 and HTTPS uses port 443; other ports, URL userinfo, and redirects to non-public destinations are rejected. URL fragments are ignored. HTTPS cannot redirect to HTTP. At most five redirects are followed, with no automatic retry or resume. Unicode hostnames use IDNA; spaces and Unicode in paths/queries are escaped without changing existing percent escapes in signed URLs.

Contract

• Formats: CSV or Parquet; gzip, bzip2, XZ, ZIP (stored or deflate members), and optional zstd wrappers are supported within resource limits
• Defaults: 64 GiB downloaded, 256 GiB expanded/decoded, 1 GiB decoded per streaming Parquet batch, and 100 million rows (tenant quotas still apply)
• Timeouts: 10 seconds to connect, 60 seconds without read progress, and 1 hour for the full transfer
• HTTP response encoding may be identity, gzip, or deflate; use an object-file wrapper for other compression formats
• Oversized, inaccessible, or expired inputs fail the job with an actionable error; a failed transfer must be resubmitted with a fresh URL
• Published Functions: columns must match Function inputs; omit column
• Direct model IDs: set column to the input column
• Optional id_column: preserved in every result format
• Use job_priority=1 and stay_attached=False

Large Parquet row groups are read in small batches. For ZIP files using bzip2 or LZMA members, recompress with deflate or use a standalone .bz2/.xz wrapper.

Sutro performs one full-object download and does not resume or retry a failed HTTP request. It can try another validated DNS address if connection setup fails. Use an immutable object and a URL lifetime that covers queueing plus download startup.

Sign an Amazon S3 object

import boto3

s3 = boto3.client("s3", region_name="us-west-2")
input_url = s3.generate_presigned_url(
    "get_object",
    Params={"Bucket": "my-bucket", "Key": "inputs/run-001.parquet"},
    ExpiresIn=3600,
)

job_id = so.infer(
    data=input_url,
    column="text",
    id_column="record_id",
    model="gpt-oss-20b",
    job_priority=1,
    stay_attached=False,
)

Cloudflare R2 and other S3-compatible stores use the same signing flow with their S3 endpoint. Google Cloud Storage uses a native V4 signed GET URL.

Treat the complete URL as a credential: do not commit or log its query string. Use a read-only signing identity, the shortest practical lifetime, and keep the source object until results are reconciled. For KMS-encrypted S3 objects, the signing identity also needs the required KMS decrypt permissions.

If pickup fails, inspect the original job before generating a fresh URL and submitting a replacement.