Why it matters
Choosing wrong wastes cluster capacity or fails user expectations. Impala for batch ETL wastes daemons; Hive for interactive dashboards times out. Understanding the trade-offs is critical.
The architecture
Impala strengths: sub-second latency, high concurrency, low resource overhead per query. Weaknesses: memory-bound (queries must fit in memory), less SQL feature coverage, less resilient to failure.
Hive strengths: complete SQL, mature UDF ecosystem, resilient to failures, handles arbitrarily large queries. Weaknesses: higher startup latency, less concurrent throughput.
How it works end to end
Impala is best for: interactive dashboards, ad-hoc analysis, BI tools with sub-second requirements, high-concurrency workloads.
Hive is best for: batch ETL, complex UDFs, very large queries that don't fit in memory, workloads that must tolerate failures.
LLAP is a middle ground: interactive-friendly Hive with cache and warm containers. Trade-off is complexity of running LLAP daemons.