Polling

The scheduler runs on a short tick, but a pool is fetched only when its own cadence says so. A block subscription moves that schedule forward whenever a block touches the pool.

A block names the accounts it changed. The subscription turns activity into an immediate poll, so a timer is unnecessary for a quiet network. The arrangement is a fallback in both directions: a dropped stream reconnects from the last processed block and the cadence keeps polls flowing while it is down, and the cadence keeps working if the stream never connects. At 10,000 dormant pools the network cost settles at about 3 requests a second instead of 667.

Registry discovery

The registry’s pools map is the pool set. The node returns at most 1,000 entries inline. Above that cap the indexer detects the limit and discovers pools through the paginated SyncAccountStorageMaps endpoint instead, which has no entry limit. It reads a block range per poll, walks up to 64 pages, and resumes from the block it reached. Discovery is idempotent, so an overlapping range costs one read and nothing else.

Manual refresh

POST /pools/{id}/refresh marks one pool due at the next tick. The response is 202 because the fetch happens on the tick, not inline. The endpoint is limited to one refresh per pool every 5 seconds and a global token bucket of 20 per second. A caller that trips either limit gets 429 with Retry-After. An unknown pool id gets 404 before either limit is consulted.

Vault reads

The vault is requested with the empty word as its asset_vault_commitment. The field is a hint: omitting it means no vault data is returned, and the response then carries an empty asset list that is indistinguishable from a genuinely empty vault. A vault the node refuses to enumerate is an error, not a zero balance.