Dynamic Checkpointing in Apache Flink

BU master's thesis. Static checkpoint intervals are a tax during idle periods and a stall during bursts. I built an adaptive controller that surfaces per-operator backpressure ratios from the Flink JobManager and uses them as a control signal: shorten cadence when load is low, lengthen under sustained backpressure to avoid amplifying stalls. Validated on the NEXMARK streaming benchmark with the RocksDB state backend; quantified write-amplification tradeoffs vs. in-memory state. Measured tail-latency wins on bursty workloads.