The customer
A B2B SaaS at Series A running ~8.4M vectors on Pinecone, paying roughly $9K/month. Their AWS commitment was 60% under-utilised, so a self-hosted move was on the table — but no one had bandwidth.
The task they submitted
Move us from Pinecone to self-hosted Weaviate. Zero customer-visible downtime. Don't break recall.
Our approach
Stood up a Weaviate cluster on their existing EKS, dual-wrote queries for 48 hours, replayed traffic to validate recall parity, then atomic-flipped the read path. Decommissioned Pinecone at day 14 once we were satisfied with the recall data.
The outcome
$9K/month → $2.7K/month. Zero downtime. Recall within 0.4% of the prior system on their internal eval set. Sixty days in production at the time of writing, no rollback.
“Migration was zero-downtime. We've been on Weaviate for 60 days and haven't looked back.”
