Workload Profiles

Vector similarity search for semantic retrieval across millions of high-dimensional embeddings. Each query computes approximate nearest neighbors against a vector index that must fit in RAM for acceptable latency. Index build is expensive; incremental updates require re-indexing strategies.

Complex analytical queries over large historical datasets, typically scanning gigabytes per query for aggregations, joins, and window functions. Query throughput is low but each query is CPU- and I/O-intensive. Freshness requirements are minutes to hours, not milliseconds.

Periodic large-volume data movement from source systems to analytical targets. High throughput over sustained windows; not latency-sensitive. Primary constraints are source read impact, destination write throughput, and job completion SLO.

Full-text search with faceted aggregations over a document corpus. Read-heavy with asynchronous write indexing. Latency-sensitive for queries; accepts index staleness of 1–30 seconds. Search relevance and facet correctness are the primary quality dimensions.

High-throughput append-only ingestion of telemetry, user events, IoT data, and clickstream records. Write throughput dominates; consumers read sequentially from the stream with bounded latency. Ordering within a partition is guaranteed; cross-partition ordering is not.

ACID-critical financial operations including ledger entries, account balance updates, and settlement. Every write must be durable, ordered, and exactly-once. Read operations for balance checks must be strongly consistent: replica reads are not acceptable for any operation that precedes a debit.

Multi-hop relationship traversal through connected entities where depth and branching factor drive query cost. Access patterns are read-heavy with point lookups to seed traversals. Query planning must bound traversal depth to prevent exponential scan growth.

High-volume transactional workload with many concurrent short-lived operations, requiring low latency, strong consistency, and high write throughput.

Multi-modal workload combining read-heavy listing search, transactional order processing, notification fan-out, and background analytics. The read path (search, browse) and write path (checkout, inventory) have very different consistency and latency requirements that must be architecturally separated.

Balanced read/write SaaS core workload covering user actions, content CRUD, and moderate read caching. Reads slightly outpace writes; the bottleneck shifts between connection pool saturation during peaks and read replica lag during write bursts.

High read throughput with infrequent writes, low latency requirements, and point-lookup or cached aggregation access patterns. Typical of content delivery, user profile APIs, and catalog services.

Sub-100ms state synchronization for collaborative editing, live cursors, presence indicators, and multiplayer game state. Ordering of operations is critical for CRDT convergence and conflict resolution. Fan-out to many connected clients amplifies write load unpredictably.

Full-text search, faceted filtering, and ranked result retrieval at high read throughput. Index refresh latency (1-second default in Elasticsearch) means freshness is near-real-time but not instantaneous. Read scaling through shard replicas is the primary lever.

Timestamped metric, event, and sensor data with extremely high write throughput and sequential read patterns (range scans over time windows). Data is append-only by nature; updates are rare. Compression ratios of 10-50x are achievable with columnar or delta-encoded storage.

High-volume durable write workload requiring ACID guarantees across order processing, payment ingestion, and audit logging. Write throughput drives system design: WAL pressure, lock contention, and index maintenance are the primary operational constraints.