A distributed snapshot must account for records between operators

A consistent cut does not include receipt of a message while excluding its send. Messages sent before the cut but received after it belong to in-flight channel state. The original Chandy–Lamport algorithm records local process states and relevant channel contents under its reliable FIFO-channel model; marker rules identify the cut without stopping all computation globally. The assumptions matter. A mechanism designed for ordered channels cannot simply be copied onto arbitrary reordered or lossy communication without additional rules. A snapshot is a logical boundary, not necessarily a set of states that all existed at one wall-clock instant.

A useful conservation check tracks a transferred quantity across sender state, receiver state, and the channel. Recording only both endpoint balances can lose an in-flight transfer or count it twice. Stream operators have analogous contributions in input buffers, output buffers, and state updates. Recovery must restore a compatible source position and replay responsibility so every required contribution has one place. This does not claim external sinks are automatically atomic; the snapshot defines the managed computation boundary, while a visible side effect can require a separate output protocol.

In Flink's aligned checkpoint mechanism, barriers follow source progress through the dataflow. At an operator with multiple inputs, arrival of one checkpoint barrier causes that input's post-barrier records to wait while other inputs continue until their corresponding barriers arrive. The operator can then snapshot a boundary that excludes all post-barrier records and forward progress. This is not a global pause of every worker at once. A slow or backpressured input can make alignment expensive, while records behind an early barrier accumulate pressure upstream.

Unaligned checkpoints instead include relevant in-flight data in the checkpoint so barriers can overtake buffered records under the mechanism's rules. This can reduce barrier delay under backpressure but adds channel-state I/O and recovery work. It is not the same as simply skipping alignment while continuing to process all inputs and omitting channel accounting; that latter at-least-once approach can replay contributions already reflected in operator state. Name the configured mode precisely. In Flink 2.1, limitations and supported combinations belong to the versioned documentation rather than an assumption that every snapshot or savepoint behaves identically.

Observe checkpoint start delay, alignment duration, synchronous snapshot work, asynchronous upload, channel bytes, and recovery time. Reducing alignment delay can expose checkpoint-storage bandwidth as the next limit; unaligned mode cannot remove an I/O bottleneck by adding more captured data. Test failure recovery under the same backpressure that motivated the change. A completed upload alone is not proof the entire checkpoint was coordinated successfully. Restore from the completed checkpoint and verify state, source positions, and sink behavior using a trace with records deliberately crossing the cut.

The queue lab illustrates waiting inventory only. Reset, set BOTH λ=3 and μ=2, predict backlog six at t=2 seconds from the initial four, then Step four times. Reset and re-set BOTH λ=2 and μ=3; predict backlog two at t=2. Editing controls retains t and recomputes from four, not a changed-rate history. These normalized fluid backlogs are not checkpoint barriers or channel snapshots, and the separate M/M/1 value is not alignment time. Use the conservation example to reason about correctness and the queue only to discuss how much work may be waiting.