Example: the same app on Vercel. The fill runs in lambda 1, the hit is served by lambda 2. The shared cache has the entry, but lambda 2's memory has no record of who filled it (no link). On a single self-hosted server it works.
Solves: whether links exist on the most common deploy target. We should know this before shipping, not after.
Fix idea: compare three options. (a) A per-process in-memory table, which works today but only within one instance. (b) Trace context stored inside the cache entry, which works everywhere but needs a Next.js upstream metadata-slot ask. (c) Both, table now and entry metadata when upstream lands. Deliverable: a decision record with rejected options and an estimate of the same-instance hit rate on serverless.
Example: the same app on Vercel. The fill runs in lambda 1, the hit is served by lambda 2. The shared cache has the entry, but lambda 2's memory has no record of who filled it (no link). On a single self-hosted server it works.
Solves: whether links exist on the most common deploy target. We should know this before shipping, not after.
Fix idea: compare three options. (a) A per-process in-memory table, which works today but only within one instance. (b) Trace context stored inside the cache entry, which works everywhere but needs a Next.js upstream metadata-slot ask. (c) Both, table now and entry metadata when upstream lands. Deliverable: a decision record with rejected options and an estimate of the same-instance hit rate on serverless.