Compare Scenarios

Migration Step 1

Both scenarios define a migration step at this stage. ML Feature Serving Platform: triggered by 'Feature computation logic duplicated across 3+ model serving'. Analytics Data Platform: triggered by 'OLTP query p99 degrading during analytics reporting windows;'.

Migration Step 2

Both scenarios define a migration step at this stage. ML Feature Serving Platform: triggered by 'Regulatory or reproducibility requirement to re-run model pr'. Analytics Data Platform: triggered by 'Sub-minute analytics freshness SLA required; ETL scheduling '.

Migration Step 3

Both scenarios define a migration step at this stage. ML Feature Serving Platform: triggered by 'Semantic similarity search recall insufficient from BM25 ful'. Analytics Data Platform: triggered by 'Dashboard query p95 > 5s on frequently accessed aggregation '.

Feature store p99 rising from < 5ms to > 50ms immediately following a model deployment or pod scaling event; Cassandra or PostgreSQL feature store read QPS spiking 10–100x simultaneously with serving fleet restart; feature store connection pool exhaustion visible in application logs during the cold start window; Redis cache hit rate dropping to < 10% for the first 60 seconds after deployment

Tier 1: Cache Cold Start Latency Spike: Simultaneous cache cold start across all serving pods after deployment, converting sequential feature requests into a thundering herd against the backing feature store. Recommended evolution: Implement pre-warming: before a new model version receives traffic, a warm-up job runs inference requests for a representative sample of entity IDs, populating the Redis cache before the pod enters the serving fleet. Use probabilistic early cache refresh (fetch from backing store before TTL expiry when remaining TTL < 20% under high request rate) to prevent simultaneous TTL expiry on hot features. Stagger deployment rollout: deploy 10% of pods, wait for cache warm, then proceed to the next 10%. .

Model performance metrics (precision, recall, AUC) degrading without a corresponding input distribution shift; ClickHouse drift dashboard showing feature value distributions served online diverging from training population distributions; explicit skew audit (comparing offline training feature values against replayed online feature values for the same entity at the same timestamp) showing systematic differences in specific features

Tier 2: Training-Serving Skew Detection: Feature computation logic divergence between the offline training pipeline and the online serving pipeline: a feature transformation applied in training is not applied identically in serving. Recommended evolution: Enforce a single feature computation function registered per feature in a shared feature registry, executed by both the online pipeline and the offline training pipeline. The two paths must use the same code, not independently maintained implementations. Implement a skew monitoring job that computes a random sample of features using both paths and alerts on distribution divergence above a threshold. .

Kafka consumer group lag (bytes or offsets) growing for the analytics topic group; ClickHouse dashboard timestamps falling behind wall clock by > 60s; ClickHouse insert throughput < Kafka produce rate

Tier 1: Consumer Lag and Freshness Degradation: ClickHouse insert throughput insufficient for Kafka produce rate. Recommended evolution: Tune ClickHouse insert buffer size and async_insert settings; increase consumer parallelism up to the Kafka partition count; batch inserts into ClickHouse using the Buffer engine or materialized views with merge trees .

One Kafka partition offset growing significantly faster than others; one consumer instance CPU/network saturated while others are idle

Tier 2: Hot Partition and Skewed Consumer Load: Skewed partition key distribution: high-cardinality entity routing the same high-volume key to one partition. Recommended evolution: Add a secondary hash suffix to the partition key to distribute load; increase topic partition count (note: keyed ordering breaks for existing messages); re-evaluate partition key selection based on actual cardinality measurements .