Andrew Miller 2bc57c073d go: keepalive, exponential backoff, chain_id metadata, durability guarantees ...

Three related fixes that turn the go template into a client that
survives the full matrix of server restart, client restart, network
blip, half-open TCP, and long outages (hours → months) — without the
user writing a line of reconnect logic in process.go.

1. gRPC keepalive: Time=10s, Timeout=3s, PermitWithoutStream=true.
   Half-open TCP (silent server restart, resumed laptop, NAT drop)
   is detected within ~13s. Previously the OS TCP keepalive took
   ~2h to notice, leaving the client as a ghost stream while prime
   logged "no active gRPC connection" for every skipped transaction.

2. Exponential backoff with jitter on reconnect. Effective delay =
   min(max_backoff_seconds, reconnect_delay_seconds * 2^attempts)
   + random(0, reconnect_delay_seconds). The attempts counter resets
   after any session that runs healthy for 60+ seconds. Jitter
   desynchronises clients so a server restart doesn't trigger a
   thundering herd. New max_backoff_seconds config field, default 120.

3. Unified error signalling: the sender goroutine now tears down the
   stream's context when it hits a Send error. Previously only Recv
   errors triggered a reconnect — a stale stream where only Send was
   broken could sit there indefinitely.

Also: chain_id is a required config field now and goes in the
x-chain-id gRPC metadata header alongside x-api-key and
x-smart-contract-id. Prime rejects streams without it with "missing
chain ID", which was silently breaking every template-based client
until users discovered it the hard way. README documents the
durability contract so contract authors know they don't have to
reimplement any of it.

2026-04-19 21:23:47 -04:00

13 KiB

Executable File

Raw Blame History

View Raw