⭐ Now accepting 3 Early Access Design Partners for Iceberg acceleration.
Accelerate Apache Iceberg workloads by 5–20× with zero application changes. Faster metadata. Smarter pushdown. Unified caching.
Works seamlessly with
The Lakehouse acceleration layer — eliminating I/O bottlenecks with smarter metadata and data pushdown.
Real Acceleration, Not Just Caching. Laketap reduces object-store latency, eliminates repetitive Parquet decoding, and applies selective pushdown to minimize CPU & network overhead.
No need to tune each engine separately. Unifies caching for Spark, Trino, Flink, DuckDB, DataFusion — with more to come.
No SQL rewrite. No table changes. No engine forks. Simply deploy Laketap and update the connector.
Tiered design: TCache holds manifest/schema, ACache Meta stores footers/indexes, ACache Data caches row-groups with selective decode—auto promote/demote based on workload.
No Lock-In. Your data never leaves your environment. Laketap is an on-prem / VPC-side acceleration layer with full enterprise isolation.
A three-step architecture that accelerates metadata, optimizes split planning, and performs compute pushdown across cache and object storage.
Architecture Overview · Metadata, planning, and data paths in one layer
Laketap is now validating TCache/ACache with design partners on real Iceberg workloads.
Catalog-facing cache that implements the Iceberg REST Catalog API and stores manifests, schema, and snapshots to cut catalog latency and avoid repeated object-store round trips.
During planning, cached Parquet footers and indexes enable row-group pruning to emit a cache-aware scan plan with only the required splits.
At execution, cached splits run selective decode and vectorized filtering; cache hits return via Arrow Flight, while cache misses stream directly from the object store, preserving native scan semantics.
Laketap isn't just a cache; it's a smart layer that learns from your data.
Automatically adjusts cache eviction, retention, and selective decode thresholds based on real-time workload usage.
Collects table/column heat, row-group access patterns, and cross-engine hotspots to dynamically optimize caching strategies.
Predicts repetitive queries, pre-warms based on daily/hourly patterns, and creates shared hotspot replicas for cross-engine workloads.
Provides recommendations for manifest layout, partition heat analysis, and compaction strategies based on historical statistics.
Early access to new planning strategies, metadata optimizations, and pushdown extensions. Collaborate directly with the engineering team.
Deploy in your VPC with a connector swap—accelerate your Lakehouse without rewriting workloads.
