Franz · Streaming Expert in Zimac

Data in motion. Made boring under failure.

Franz follows one event from produce to side effect, crashes it at every boundary, and names the first invariant that breaks. Then he gives you the smallest move that makes the flow hold.

Run the Failure Film Read the guide

Kafka · Kinesis · queues · pub/sub · CDC · streaming ETL

01

TraceOne event. Every persist, retry, side effect, and ack.

02

ScoreDelivery, ordering, recovery, backpressure, and capacity.

03

ObserveLag, skew, ISR, schemas, connectors, and live telemetry.

04

MoveExact target and arguments. Your approval before mutation.

Franz in a streaming operations room, wearing his three-wave desktop mark while three event lanes pass through a relay behind him
FranzEvent Systems Architect ≋

Verdict → Signal → Move

Kinetic calm for systems that cannot stand still.

Franz treats an event system as a temporal contract: identity, ordering scope, delivery boundary, replay path, schema, observer, and owner. He is fluent in the broker you have—and gently contrarian when a queue is enough.

“What happens when this consumer crashes mid-batch?”
  • ObservedWhat the cluster, code, or telemetry actually says.
  • DerivedWhat deterministic capacity and lag math supports.
  • UnknownWhat has not been measured yet—without false precision.

Franz's signature move

Crash the happy path. Keep the contract.

The Failure Film follows one event across every irreversible boundary. Choose a cut: Franz shows the counterexample, the violated invariant, and the smallest repair.

01Produceevent id assigned
02Persistbroker confirms
03Consumebatch received
04Side effectcharge succeeds
05Commitoffset pending
VERDICTDuplicate side effect

The charge succeeds, the consumer dies before committing, and the event is delivered again.

SMALLEST MOVE

Make the side effect idempotent on the event identity, then commit only after the durable result.

Deterministic design review

A score with a reason to move.

Franz scores the design you actually describe—not a generic Kafka checklist. The rubric covers delivery, ordering, recovery, backpressure, partitioning, capacity, schema evolution, and lag observability.

  • ≋Broker-aware. SQS is judged as SQS, not Kafka with fewer knobs.
  • ≋Failure-first. The first broken invariant outranks the longest risk list.
  • ≋Exactly-once skeptical. The side effect boundary must prove the claim.
STREAM REVIEW · CHECKOUT EVENTSillustrative
72/100CONDITIONAL
FIRST BROKEN INVARIANT

A replay can charge twice because event identity stops at the consumer boundary.

Capacity reality check

Throughput in. Recovery truth out.

Give Franz ingress, payload size, partitions, retention, and measured consumer rate. His computed card keeps every assumption visible and refuses to invent a time estimate when the rate is missing.

PEAK INGRESS24.6 MiB/sDerived
MIN PARTITIONS18Derived
6H OUTAGE BACKLOG519 GiBDerived
CATCH-UP ETA47 minRate observed

ASSUMPTIONS 2.0× peak factor · replication factor 3 · 30% headroom · 7-day retention

Read the live pulse

Evidence before diagnosis.

Connect Kafka, MSK, Kinesis, Datadog, Grafana, New Relic, Splunk, GitHub, or GitLab. Franz starts with the narrow read that answers the question and labels where every signal came from.

KafkaMSKKinesisConnectSchema RegistryJMX
payments-prod
OBSERVED · 14:32:08 UTC
GROUPcheckout-workers
STATEDEGRADED
Total lag48,210+1,840/min
Worst partition17,904p_07 · 37%
Under replicated0healthy
SIGNAL

Input is steady. One partition owns 37% of lag after a consumer joined. Inspect key distribution before adding capacity.

FRANZ REQUESTS APPROVALWRITE
OPERATIONconsumer_group.offsets.reset
Cluster
payments-prod
Group
checkout-replay
Topic
checkout.events
Target
2026-08-07T18:00:00Z
BOUND REQUEST8be2…a914
Illustrative · changed arguments require a fresh decision

Read first · guard every write

Franz cannot approve himself.

Topic changes, offset resets, connector restarts, schema updates, rebalances, and MSK or Kinesis mutations pause on an exact-target approval card. Your one-time click resumes that request—and only that request.

01

Resolved targetCluster, resource, operation, and arguments are visible.

02

Hash-boundChange the request and the old approval no longer applies.

03

One useA completed approval cannot be replayed.

Gently contrarian by design

Sometimes a queue is enough.

Franz knows Kafka, Flink, Kinesis, Pulsar, SQS, SNS, EventBridge, Pub/Sub, RabbitMQ, NATS, CDC, and streaming ETL. Expertise is knowing the whole toolbox. Judgment is not reaching for all of it.

Choose the smallest system whose failure contract you can actually operate.

Franz lives inside Zimac Pro

Give him a flow, a failure, or a live cluster.

Start with a pasted design. Connect evidence when the question reaches production.

Get Zimac Explore Franz's guide