One Writer, Fifty Readers: How Far a Single-Primary Database Actually Goes

Subtitle: One writer. Fifty readers. All load on one box.

• Panel one - The one box that cannot be duplicated:
• 1 primary takes every write
• ~50 read replicas across regions
• replicas are copies fed by the WAL stream
• the goal: keep work OFF this box

• Panel two - Five defenses, one move each:
• offload reads to the replicas
• pool connections: 50ms down to 5ms
• isolate noisy workloads
• lock the cache on miss
• shed load by query digest

• Panel three - What the headline hides:
• MVCC copies the whole row per update
• dead tuples, bloat, vacuum chores
• each replica adds WAL load to the writer
• it scales = read-heavy + writes moved out

• Panel four - Falsify 'reads are offloaded':
• instrument the connection layer
• count read-only queries on the primary
• the leak is usually ORM reads in a write tx
• near zero? else the outage is still total