Impala Cluster Performance Optimization Based on Historical Queries: Practices and Solutions

01 Background Introduction

NetEase DataFun is a digital transformation technology and service provider under NetEase, comprising NetEase Lightboat, NetEase YouShu, NetEase YiZhi, and NetEase YiCe. YouShu supports most internal internet services (music, selection, YouDao, email, media) and many external customers across finance, retail, agriculture, energy, logistics, etc.

Impala in YouShu focuses on interactive computing, serving BI, self‑service data extraction, and ad‑hoc queries.

High‑Performance Data Warehouse Construction

The NetEase big‑data team was among the first to adopt Impala as an analytical engine. Over years they accumulated extensive experience in cluster operation, troubleshooting, feature enhancement, and performance tuning. Pure community Impala proved insufficient; without further enhancements, query performance and success rates suffered.

Key pain points identified:

Mixed workload challenges causing resource contention and SLA violations.

Poor performance of complex/aggregated queries, especially BI reports requiring sub‑5‑second response times.

Missing statistics leading to inaccurate cost estimates and slow queries.

Metadata cache staleness causing errors when underlying tables change.

Storage layer fluctuations (HMS/HDFS) affecting query latency.

Periodic cluster state degradation due to increasing workload and resource competition.

To address these, the team proposes a set of construction principles for a high‑performance data warehouse, emphasizing continuous enhancement of Impala’s core capabilities and leveraging historical query information.

03 HBO Implementation Plan

HBO (Historical‑Based Optimization) collects query SQL and profile data via the Impala Manager component, persisting them in MySQL. The Manager registers with the Statestore, retrieves query metadata from coordinators, and stores detailed information such as queue time, memory estimates, actual consumption, scanned data volume, profile, timeline, and summary.

Analysis of these records enables identification of performance bottlenecks, memory‑estimate correction, and detection of metadata cache issues.

Key implementation steps:

Enable virtual warehouses to isolate mixed workloads across physical resources.

Automate statistics collection and use historical query memory consumption to refine estimates.

Enhance Impala’s local cache (metadata and HDFS file cache) and add Footer Cache for Parquet files.

Introduce multi‑table materialized views using Calcite for automatic SQL rewriting, supporting outer joins, group‑by, limit, and custom UDFs.

Provide metrics for materialized‑view hit rates, rewrite cost, and storage usage.

Improve memory‑estimate accuracy via template‑based adjustments.

Deploy virtual warehouses via Zookeeper namespaces or request‑group routing for fine‑grained load balancing.

04 Future Development Plans

Roadmap includes:

Impala 4.1 release with merged self‑developed features and Hive 2.x support.

Expanded materialized‑view capabilities, higher automation, and tighter integration with YouShu BI.

Kernel optimizations: better memory estimation, transparent retry, vectorized execution.

Cluster management enhancements: K8s‑based deployment, dynamic resource scheduling, health‑diagnosis system.

The session concludes with thanks to the audience.