How SnapshotScanMR Doubles HBase‑to‑Hive ETL Speed and Relieves Cluster Load

Background

In the Zhongtong business scenario, massive back‑fill of 7‑15‑day data is required. Performing this entirely offline would overload the business system, so HBase rowkey updates are used to store historical data. Daily ETL jobs must pull billions of rows from HBase, scanning close to a hundred billion records.

Problems with the Old Solution

The traditional approach used HBaseStorageHandler to map HBase tables to Hive, extracting data into a new Hive table. This caused a huge number of requests to HBase region servers, triggering load alerts, especially during nightly peak periods, and adversely affecting other cluster tasks.

Investigation of HBase and Hive source code revealed no existing support or optimization for this workload.

New Solution: HBase‑to‑Hive via SnapshotScanMR

By exploiting the low‑level SnapshotScanMR feature, a new task type hbase2hiveBySnapshot was developed on the big‑data platform, completely eliminating the performance issues of the previous method.

The task first creates a snapshot of the source HBase table, parses user‑provided conditions (filters, table name, etc.), and then uses a custom InputFormat that treats each HRegion's HFile as a map input. Reduce tasks are sized according to table size, filter data according to user criteria, and finally write the results to HDFS, which are then loaded into the corresponding Hive table/partition.

Testing shows that the new task imposes zero pressure on HBase region servers, cuts execution time by about 50%, reduces resource usage, and significantly improves cluster stability.

Why SnapshotScanMR Solves the Issue

Unlike the original task that scans the live table via TableScanMR —sending many limited next requests to region servers—SnapshotScanMR reads the snapshot (restored HFiles) directly from HDFS on the client side, bypassing region servers entirely. This dramatically reduces the number of requests to region servers and doubles scanning efficiency.

Performance Gains

Benchmark results indicate a 2× improvement in scan speed and a complete elimination of region server load.

Future Improvements

Planned enhancements include adding filter support, skipping empty HFiles, enabling more flexible task splitting (custom region partitioning), and gathering community feedback for further optimization.