Skip to lesson
Exit
Lakehouse & Ingestion at Scale1 / 3

1 min lesson

SLAs and how you'd alert on each

Choose two examples from the table in "SLAs and how you'd alert on each" and explain what each teaches you to do.

Step 1 of 3

SLAs and how you'd alert on eachfreshness and completeness are the two that matter

SLA
Freshness
Definition
Silver lags bronze by < N minutes
Alert signal
Max event-time watermark vs now exceeds threshold
SLA
Completeness
Definition
Daily event count within expected band
Alert signal
Volume anomaly vs trailing baseline (per source)
SLA
Lag
Definition
Consumer keeps up with the stream
Alert signal
Kafka consumer lag rising / not draining
SLA
Schema health
Definition
No rejected/quarantined spike
Alert signal
Dead-letter rate above baseline

Tie every SLA to a concrete metric you can alert on - vague 'monitor it' answers don't land.

Learn more

Full explanation

Failure modes and recovery

Failure modes and recoverythe part that proves you've run one of these

Late / duplicate events

Watermarks bound how late you'll wait in silver

Idempotent MERGE on event id collapses duplicates

Bronze retains raw, so a wider replay is always possible

Schema drift

Registry rejects breaking changes at the producer

Bronze permissive, silver strict - one table degrades, not all

Quarantine bad rows, alert, fix, replay from bronze

Consumer lag / backpressure

Stream buffers the spike; transforms drain at their pace

Autoscale consumers up to the partition ceiling

Alert when lag stops draining, not on every blip

Downstream outage

Ingest keeps landing bronze independently

Backfill silver/gold from bronze once recovered

No data lost because raw is durable and replayable

Say it like this

“Client events hit a thin ingest gateway that stamps a server timestamp and an event id, then publish to Kafka topics partitioned for our peak - Kafka is the durable buffer so a client spike never touches downstream compute. Structured Streaming appends to bronze Delta, partitioned by event date, with an idempotent MERGE on the event id so replays don't double-count. Bronze→silver dedups, enforces the registered schema and uses watermarks for late events; malformed rows go to a dead-letter topic I can replay after a fix. Gold and the agent feature tables are derived from silver and fully rebuildable. I'd SLA on freshness and completeness, alerting on watermark lag and a volume anomaly per source.”

Interview move

Close the design by naming what you'd cut for v1 and what you'd add at 10x. "For v1 I'd ship gateway → Kafka → bronze → silver with idempotent writes and a dead-letter path and defer liquid clustering and the feature store until the access patterns are real." Volunteering the staging plan signals the pragmatic, ship-fast judgment the role is filtering for, not gold-plating.