We describe extensions to Confluo's interface to capture time-series data and support operations similar to BTrDB on the captured data.
Time-series data comprises of a stream of records, each of which is a (timestamp, value) pair. Confluo maintains an index on both the timestamp and the value attribute on the Atomic MultiLog to support queries on time windows as well as more general diagnostic queries. Confluo also supports efficient aggregate queries on the captured time-series data, but via pre-defined aggregates and ad-hoc query execution.
Atomic MultiLog offsets implicitly form versions for the timeseries database — each offset corresponds to a new version of the database and includes all records before that offset. Confluo also permits users to compute the difference between two versions of the database: since database versions map to Atomic MultiLog offsets, Confluo fetches all records that lie between the offsets corresponding to the two versions.
Compared Systems and Experimental Setup¶
We evaluate Confluo against BTrDB, CorfuDB and TimescaleDB on c4.8xlarge instances with 18 CPU cores and 60GB RAM. We used the Open μPMU Dataset, a real- world trace of voltage, current and phase readings col- lected from LBNL’s power-grid over a 3-month period. We create a separate Atomic MultiLog for each type of reading (voltage, current or phase). We run single server benchmarks with 500 million records to highlight the per-server performance of these systems. Requests are issued as continuous streams with 8K record batches.
Figure: ConfluoD & ConfluoR measure performance for DURABLE & DURABLE_RELAXED modes respectively. ConfluoD achieves 2-20x higher throughput, 2-10x lower latency for inserts, and 1.5-5x higher throughput, 5-20x lower latency for time-range queries than compared systems.
The figure above shows that systems like CorfuDB and TimescaleDB achieve over 10x lower performance than BTrDB and Confluo. We emphasize that this is not a shortcoming: CorfuDB and TimescaleDB support stronger (transactional) semantics than BTrDB and Confluo. Thus, depending on desired semantics, either class of systems may be useful for an underlying application.
Confluo, using DURABLE_RELAXED writes, is able to achieve close to 27 million inserts/second due to its cheap versioning and lock-free concurrency control. For time-range queries, almost all systems observe similar throughput since queries are served via in-memory indexes. Insertion and query latency trends are similar to throughput trends across different systems.