Jun Qin and Nico Kruber have started a series on low-latency streaming in Apache Flink. The first two posts of the series are up, starting with the overview:
Latency can refer to different things. LatencyMarkers in Flink measure the time it takes for the markers to travel from each source operator to each downstream operator. As LatencyMarkers bypass user functions in operators, the measured latencies do not reflect the entire end-to-end latency but only a part of it. Flink also supports tracking the state access latency, which measures the response latency when state is read/written. One can also manually measure the time taken by some operators, or get this data with profilers. However, what users usually care about is the end-to-end latency, including the time spent in user-defined functions, in the stream processing framework, and when state is accessed. End-to-end latency is what we will focus on.
Part 2 discusses direct latency optimization techniques:
When interacting with external systems (e.g., RDBMS, object stores, web services) in a Flink job for data enrichment, the latency in getting responses from external systems often dominates the overall latency of the job. With Flink’s Async I/O API (e.g., AsyncDataStream.unorderedWait() or AsyncDataStream.orderedWait()), a single parallel function instance can handle many requests concurrently and receive responses asynchronously. This reduces latencies because the waiting time for responses is amortized over multiple requests.
Stay tuned for more posts in the series.