fixed-resolution-sidecar

name	fixed-resolution-sidecar
description	Build 512-area and 1024-area resized image Parquet datasets, preserve row keys, audit valid/invalid rows, and feed crop metadata back into structured_description bbox columns.

Fixed Resolution Sidecar

Purpose

Build resized image Parquet datasets from a canonical main Parquet without rewriting the original dataset. The common Core CN outputs are:

• 0_262mp_version_all_in_one/<dataset>/ for 512-area images.
• 1_05mp_version_all_in_one/<dataset>/ for 1024-area images.

These are image Parquet datasets. They contain resized image bytes and audit metadata. The final all-in-one sidecar later uses their resize metadata to create structured_description_crop_512, structured_description_crop_1024, and bbox_resize_meta.

Output Contract

For each resized output:

• Preserve the same relative shard layout as the main Parquet.
• Preserve the row order and key column.
• Preserve aspect ratio. Do not force a square crop unless explicitly required.
• Use WEBP quality 95 unless there is a dataset-specific reason to change it.
• Keep invalid rows with audit metadata when the implementation supports it; do not silently drop rows.

Recommended prefixes:

512  -> s3://core-cn-oss.canva.com/dataset/0_262mp_version_all_in_one/<dataset>/
1024 -> s3://core-cn-oss.canva.com/dataset/1_05mp_version_all_in_one/<dataset>/

UTP Pattern

Use sidecar-ops build-0-262mp or the current resize op in sidecar_parquet_ops. Keep jobs CPU-only.

Example shape:

name: ${arnold_job_name}
max_retries: 1

image_uri: "${arnold_docker_build:ds-core-cn-omra-trainer,.,deploy/Dockerfile.resize_cpu,deploy_core-cn-readwrite}"

entrypoint: >
  sidecar-ops build-0-262mp
  --main-prefix s3://core-cn-oss.canva.com/dataset/<dataset>_parquet/
  --output-prefix s3://core-cn-oss.canva.com/dataset/0_262mp_version_all_in_one/<dataset>/
  --key-column image_key
  --audit-prefix 0_26mp_process
  --patch-size 32
  --use-ray
  --skip-existing
  --max-inflight 32
  --task-cpus 4
  --task-memory-gb 12
  --threads-per-shard 4
  --output-format webp
  --quality 95

env_vars:
  CANVA_FLAVOR: prod
  FLAVOR: prod
  CANVA_PLATFORM: eks
  PIP_CONSTRAINT: ""
  RAY_SCHEDULER_EVENTS: '0'

compute_config:
  head_node:
    instance_type: r5.4xlarge
  worker_nodes:
    - name: cpu_worker
      instance_type: m5.4xlarge
      min_nodes: 8
      max_nodes: 8
      market_type: 'ON_DEMAND'
  flags:
    workload_starting_timeout: 1h

For 1024-area output, use the repo's current 1024 resize command or pipeline variant and set:

--output-prefix s3://core-cn-oss.canva.com/dataset/1_05mp_version_all_in_one/<dataset>/
--audit-prefix 1_05mp_process

Crop Bbox Merge

After both resized image Parquet outputs exist, run a sidecar pipeline that reads:

• original structured_description
• 512 resize audit metadata
• 1024 resize audit metadata

and writes:

• structured_description_crop_512
• structured_description_crop_1024
• bbox_resize_meta

The original structured_description stays in the original image coordinate frame. Crop variants must also use normalized 0-1000 yxyx bboxes.

Verifier

Run these checks before using the resized outputs downstream:

1. Shard count:

   AWS_PROFILE="${AWS_PROFILE_NAME}" aws s3 ls s3://.../main/ --recursive --summarize
   AWS_PROFILE="${AWS_PROFILE_NAME}" aws s3 ls s3://.../0_262mp_version_all_in_one/<dataset>/ --recursive --summarize
   AWS_PROFILE="${AWS_PROFILE_NAME}" aws s3 ls s3://.../1_05mp_version_all_in_one/<dataset>/ --recursive --summarize
   

2. Sample key alignment and decode:

   • same row count as source shard
   • same key column values
   • image bytes decode with PIL
   • dimensions are near target area and preserve aspect ratio

3. Audit columns:

   • count valid and invalid rows
   • report invalid examples
   • verify invalid rows are not silently omitted

4. Crop caption columns:

   • structured_description_crop_512 exists and has valid bboxes
   • structured_description_crop_1024 exists and has valid bboxes
   • bbox_resize_meta records source and target dimensions
   • sampled overlay visualization renders original, 512, and 1024 bboxes

Guardrails

• Treat "512" and "1024" as fixed image area unless the project explicitly requires square crops.
• Never overwrite the main Parquet.
• Keep resized outputs separate from the all-in-one sidecar.
• Use --skip-existing when available.
• Do not proceed to crop-aware JSON caption merge until both resized outputs have verifier results.