Gateway cache warming

[vasco-santos]

Gateway warming cache proposal

Anecdotal evidence suggests that recent uploads have a high probability to be requested from a gateway in a near future.

The solution proposed here aims to improve the performance of reads on recently uploaded content by warming the cache immediatly after write. Indirectly, it will also help decrease load on public gateways and work around delays on data availability.

High level flow

1. Write backup upload chunk to S3 bucket nft.storage/packages/api ✔️
2. Lambda function triggers on S3 write assemble-cars-lambda
   • verify if /raw namespace trigger
   • perform CAR completeness optimistic validations
     • is CAR file complete? we can rely on S3 metadata
   • if CAR is complete: write to /complete namespace
   • else: does CAR file have a DagPB root with know acceptable size?
     • get list of CARs in same directory and verify if their total size is bigger than known size
     • Get all CARs in the S3 directory
     • Validate CAR is complete through all the links
     • Join CARs and write to /complete namespace
3. Lambda function triggers on S3 write gateway-warm-cache-lambda
   • verify if /complete namespace trigger
   • ask Gateway worker to cache CID
4. CF Gateway Worker Receives a request to look for a CID and cache it nft.storage/packages/gateway
   • Asks S3 Gateway for CID content
   • Put response in CF cache
5. S3 exporter s3-exporter or s3-gateway
   • Read requested CID from S3 /complete namespace
   • Unpack CAR, unixfs export it and stream it

Implementation details

• assemble-cars-lambda
  • We need to look into limitations of AWS Lambda memory for "acceptable" size config
• gateway-warm-cache-lambda
  • Only ask to warm cache for content smaller than 512Mb
• Gateway Worker
  • Route /cache/:cid
    • Protected route! Only our internal services should be able to do this!
  • We can use S3 exporter as a gateway in the race
• S3 exporter
  • https://github.com/ipfs/js-ipfs/blob/master/packages/ipfs-http-gateway/src/resources/gateway.js

Metrics

• Successful warm cache requests
• Length distribution of warm cache content

How future with stream uploads looks like

TBD Notes:

• With stream uploads, we will be able to know in the API when uploads are wrapped up
  • (Probably) We will be able to drop the lambda function and API Worker can just ask Gateway Worker to warm cache when upload ends -- Cache content needs to be added within gateway worker region ID

Alerting system

Users have reported intermittent issues trying to get data from public gateways.

We have been considering the read test lambda as way of creating alerts but that will result in a lot of extra load to the gateways, which at least for now would not be desirable.

TBD Notes:

• @alanshaw https://github.com/nftstorage/checkup
• We can randomly try a few of the finalized uploads directly to ipfs.io gateway and track metrics with Lambda function logs

This will drop tracked work on:

• https://github.com/nftstorage/nft.storage/issues/1405
• https://github.com/nftstorage/nft.storage/issues/755