Zone-aware consumer reads for JetStream (similar to Kafka's KIP-392)

[bobo]

Hello!

When running JetStream with R3 across availability zones, the same message payload can cross AZ
boundaries up to three times — Raft stream replication, consumer leader delivering to clients, and core
NATS routing to subscribers. At high throughput that adds up.

The data is already on the local replica. Kafka addressed this with KIP-392, letting consumers fetch
from the nearest broker holding a replica instead of always going to the partition leader.

JetStream already has a precedent — AllowDirect lets all stream replicas serve Direct Get requests,
trading read-after-write consistency for read distribution. For pull consumers, something similar could
work:

• Client indicates topology preference (perhaps via existing server tags or a new consumer config field)
• Local stream replica serves the message payload
• Consumer state (delivery tracking, acks) still flows through the consumer leader via Raft
• Consistency stays at-least-once — AckExplicit already handles redelivery on leader failover

Is this something the team has considered? Are there architectural reasons it would be harder than it
sounds? Would mirrors be a better path than in-cluster replicas?

Happy to dig deeper into any aspect of this.

[derekcollison]

Yes we have been exploring and looking at ways to optimize costs for our customers for both stream ingest and consumption as you mentioned above. Nothing definite yet, but we are researching ideas.

[bobo]

Sounds good!
Is there anything we could contribute with? This is fairly important for us.
If i where to draft a more concrete proposal building on my message above, is that something that would be interesting to add? I can ofc also look into implementation.

[derekcollison]

Best way to contribute is to be a customer of Synadia.