roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepare for Apache Flink interview questions on how streaming jobs handle time, state and failure. Covers the DataStream API and Flink SQL, event time and windows, state and checkpoints, and deployment.
49 of 49 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 49 new. 0 cut, outside your plan.
Flink is rarely the subject of a whole loop. It arrives as one round inside a data, backend, platform or analytics interview, and which team is hiring decides what that round asks. Data engineers on a streaming platform get the correctness questions: why an hourly count comes out different when the job replays yesterday's data, what happens to an event that arrives after its window has closed, and how a job reading from Kafka and writing back to it avoids emitting a result twice after a restart. Backend teams that build fraud checks, alerts or live dashboards on Flink look at the code in a `KeyedProcessFunction`: what belongs in keyed state, when a timer fires, and what that state costs once the key space is large. Platform teams that run Flink for other people go straight to operations: backpressure, checkpoints that take longer every hour, and a savepoint that will not restore because the job graph changed and its operators had no explicit `uid`. Teams that write only Flink SQL ask why a plain join between two streams keeps both sides in state indefinitely, and what a state time-to-live gives up to stop it. The round usually sits inside a wider conversation. A data engineering interview reaches well past streaming, and the [Data Engineer roadmap](/roadmaps/data-engineer) prepares those other questions; most Flink jobs read from a log such as Kafka, and the [Kafka roadmap](/roadmaps/kafka) covers that side of the pipeline; and interviewers who frame Flink against Spark Structured Streaming expect you to know that engine as well, which the [Spark roadmap](/roadmaps/spark) prepares.
Read the [primer on the Flink hub](/topics/data-flink) once, for how time, state and failure depend on one another, then use this page to drill. The primer earns a second visit whenever an answer you give about one part turns out to rest on another part you cannot yet explain. A junior gets [event time and watermarks](/topics/data-flink-event-time-watermarks) solid first, on top of keyed streams: what a watermark asserts, why the slowest input holds it back, and what a window does with an event that arrives behind it. A middle-level round spends its time on [state](/topics/data-flink-state-management): which state type fits a per-key deduplication or a running session, where the state backend keeps it, and why it keeps growing when nothing clears it. Seniors are pressed hardest on [checkpointing and exactly-once](/topics/data-flink-checkpointing-exactly-once), and the questions come as incidents: barriers that cannot align under backpressure, a checkpoint that times out, a sink whose output stays invisible until the next checkpoint commits it. If you already know Spark Structured Streaming, keep the junior band in the level range anyway. Both engines speak of watermarks and windows, the details differ, and a Spark-shaped answer to a definitional question is easy to give without noticing.
Flink is rarely the subject of a whole loop. It arrives as one round inside a data, backend, platform or analytics interview, and which team is hiring decides what that round asks. Data engineers on a streaming platform get the correctness questions: why an hourly count comes out different when the job replays yesterday's data, what happens to an event that arrives after its window has closed, and how a job reading from Kafka and writing back to it avoids emitting a result twice after a restart. Backend teams that build fraud checks, alerts or live dashboards on Flink look at the code in a KeyedProcessFunction: what belongs in keyed state, when a timer fires, and what that state costs once the key space is large.
Platform teams that run Flink for other people go straight to operations: backpressure, checkpoints that take longer every hour, and a savepoint that will not restore because the job graph changed and its operators had no explicit uid. Teams that write only Flink SQL ask why a plain join between two streams keeps both sides in state indefinitely, and what a state time-to-live gives up to stop it.
The round usually sits inside a wider conversation. A data engineering interview reaches well past streaming, and the Data Engineer roadmap prepares those other questions; most Flink jobs read from a log such as Kafka, and the Kafka roadmap covers that side of the pipeline; and interviewers who frame Flink against Spark Structured Streaming expect you to know that engine as well, which the Spark roadmap prepares.
Read the primer on the Flink hub once, for how time, state and failure depend on one another, then use this page to drill. The primer earns a second visit whenever an answer you give about one part turns out to rest on another part you cannot yet explain.
A junior gets event time and watermarks solid first, on top of keyed streams: what a watermark asserts, why the slowest input holds it back, and what a window does with an event that arrives behind it. A middle-level round spends its time on state: which state type fits a per-key deduplication or a running session, where the state backend keeps it, and why it keeps growing when nothing clears it. Seniors are pressed hardest on checkpointing and exactly-once, and the questions come as incidents: barriers that cannot align under backpressure, a checkpoint that times out, a sink whose output stays invisible until the next checkpoint commits it.
If you already know Spark Structured Streaming, keep the junior band in the level range anyway. Both engines speak of watermarks and windows, the details differ, and a Spark-shaped answer to a definitional question is easy to give without noticing.