Ordering & FIFO

Ordering and concurrency are fundamentally at odds: if message B must wait for message A, no amount of workers makes B faster. Systems that pretend otherwise either serialize everything (slow) or order nothing (wrong). Vulkan’s answer is the same as Kafka’s and SQS FIFO’s — order within a key — but opt-in per message rather than baked into stream topology.

Partition keys

Tag messages that must be ordered relative to each other with a key:

// All events for one account process in order…
stream.Publish(ctx, &BalanceAdjusted{Account: "acct-42", Delta: -500},
    vulkan.WithPartitionKey("acct-42"))

// …while unkeyed messages parallelize across every worker.
stream.Publish(ctx, &EmailRequested{To: "x@example.com"})

The guarantee: within a consumer group, at most one message per partition key is in flight at a time, claimed in log order. Messages with different keys — and messages with no key — run fully parallel. Fifty workers can chew through a stream while acct-42’s events march through single-file.

Under the hood it’s one predicate in the claim query: skip rows whose key already has an in-flight delivery in this group. No partition count to choose up front, no rebalancing, no “we picked 12 partitions in 2024 and now we’re stuck with them.”

The hot-key trade, made visible

A hot key serializes itself — that’s not a bug, it’s what you asked for. If one account produces 80% of your traffic, that account processes at single-worker speed while everything else saturates the pool:

-- Who's hot right now?
SELECT e.partition_key, count(*) AS backlog
FROM vulkan.deliveries d
JOIN vulkan.events e ON e."offset" = d.event_offset
WHERE d.status = 'ready' AND e.partition_key IS NOT NULL
GROUP BY 1 ORDER BY 2 DESC LIMIT 10;

(Vulkan Cloud charts this and alerts when a key starts dragging group latency.)

Compared to the neighbors

	Kafka	SQS FIFO	Vulkan
Ordering unit	partition (fixed count)	MessageGroupId	partition key
Choose order per message?	❌ topic-wide topology	✅	✅
Unordered fast path on same stream	❌	❌ separate queue	✅ null key
Reshard pain	repartitioning project	n/a	none — keys are just values