Scala engineer behavioural interview questions

As asked

Tell me about a time you drove a large-scale migration or upgrade (language version, library, or framework) across multiple teams. What was your approach to getting buy-in, and how did you handle teams that were reluctant to change?

Sample answer outline

A strong answer uses the STAR format with a concrete migration (Scala 2 to 3, Akka 2.5 to 2.6+, or a major library change). The situation should include cross-team coordination, competing priorities, and resistance. The action includes building a proof of concept, publishing migration docs, pairing with reluctant teams, and using feature flags or compatibility layers to reduce risk. The result includes measurable outcomes like reduced compile times or fewer runtime errors.

Expect these follow-ups

• How did you handle a team that missed their migration deadline?
• What would you do differently the next time you run a large migration?

As asked

Tell me about a time a Spark job passed all staging tests but failed or produced wrong results in production. Walk me through how you debugged it.

Sample answer outline

Look for differences between staging and production: data volume (skew that did not appear at small scale), schema differences (production has a column staging lacked), date range issues (full history versus recent only), or config differences (partition count, memory). The action should include checking Spark UI stage metrics, adding logging to identify which task failed, and adding a data quality step that catches the discrepancy early. Strong answers distinguish between a correctness bug and a performance failure.

Expect these follow-ups

• How do you write Spark jobs that are easier to test in a local mode?
• What monitoring would have caught this failure before it reached production?

As asked

Describe a time you had to investigate a sudden latency increase in a live service. What was your systematic approach, and how did you identify the root cause?

Sample answer outline

Start with the observation: what changed (deployment, traffic pattern, upstream). Then narrow with metrics: CPU, memory, GC activity, actor mailbox depth, thread pool saturation. Use async-profiler or JFR for a flame graph. Common Akka causes: blocking call on the actor dispatcher, mailbox overflow, or a slow ask with short timeout cascading failures. The result should be a specific root cause and a fix with measurable before/after metrics.

Expect these follow-ups

• How do you distinguish a JVM GC pause from an application logic slowdown in your metrics?
• What alerts would you set up to catch this problem earlier next time?

As asked

Tell me about a time you disagreed with a significant technical decision in a Scala project, for example choosing Akka actors over a pure functional effect system, or staying on Scala 2 when you believed migrating to Scala 3 was the right call. What did you do and what was the outcome?

Sample answer outline

A strong answer shows the candidate raised the concern proactively with data (benchmark, proof of concept, documented trade-offs) rather than just an opinion. They should show they were willing to disagree but also commit once the decision was made. The outcome can be that they changed the decision with evidence, or that they did not but the decision still worked out (or taught them something). Avoid stories where they just silently complied or fought indefinitely.

Expect these follow-ups

• What would you do if you were overruled on a decision you still believe is wrong?
• How do you make sure your disagreement does not damage your working relationship with the tech lead?

As asked

Describe a time you were under a hard deadline on a Scala service and had to decide between shipping quickly with a shortcut, such as using a mutable var or throwing exceptions instead of typed errors, versus taking more time to do it the functional way. What did you decide and how do you feel about it now?

Sample answer outline

The answer should demonstrate the candidate's decision framework: what constitutes an acceptable shortcut (skipping a non-critical feature) versus an unacceptable one (skipping error handling on financial transactions). They should have created a tech-debt ticket and addressed it afterward. Strong answers show the candidate made the trade-off explicitly with stakeholders, not unilaterally, and can reflect honestly on whether the shortcut was the right call.

Expect these follow-ups

• How do you ensure technical debt from deadline pressure actually gets addressed afterward?
• What is a shortcut you have taken that you regret?

As asked

Tell me about a time you were on call and discovered that a data pipeline had silently lost or duplicated data. How did you respond, and what did you change afterward to prevent it?

Sample answer outline

The story should cover immediate containment (stop the pipeline, assess scope of loss), root cause investigation (missing deduplication, offset reset, schema mismatch), and communication to stakeholders. The follow-up changes should be concrete: idempotency keys, row count checks, offset tracking in the database, or alerting on unexpected output record count drops. A strong answer quantifies the impact and shows what monitoring gap was closed.

Expect these follow-ups

• How do you communicate data loss to non-technical stakeholders?
• What is your checklist for validating a data pipeline before enabling it for production traffic?

As asked

Tell me about a time you reviewed code from a colleague who came from an imperative background and wrote Scala in a way that missed functional programming principles. How did you give feedback and help them improve?

Sample answer outline

The story should include a specific antipattern (mutable var in a for loop, throwing exceptions instead of returning Either, Future.get blocking a thread). The feedback approach matters: specific, educational, focused on the why rather than just the what. A strong answer includes offering to pair-program, pointing to team style guides, and following up to see if the pattern changed in subsequent PRs.

Expect these follow-ups

• How do you balance functional purity against the pragmatic need to ship features?
• What team-level practices (linters, style guides) do you advocate for to reduce this kind of feedback?

As asked

Tell me about onboarding a new engineer to a complex Scala codebase that used effect systems, implicits, or type class hierarchies. How did you help them get productive and what misconceptions did you have to unpack first?

Sample answer outline

A strong answer shows structured onboarding: architectural overview, guided code reading, pairing on a small real task, and deliberate documentation of decisions. The measurement can be time to first PR, time to first independent task, or a direct check-in with the new engineer. A good candidate also reflects on what they learned about the codebase itself by explaining it to someone new.

Expect these follow-ups

• What is the most common misconception a new Scala engineer has about functional programming in practice?
• What documentation do you wish existed when you joined your current team?

As asked

You join a team and discover the existing Scala service throws raw exceptions everywhere instead of using Either or a typed error channel. The codebase is large and actively being developed. What do you do?

Sample answer outline

Avoid a big-bang rewrite. Instead, introduce a boundary: new code written in the functional style, and adapters at the boundary that catch exceptions and convert them to Either or IO. Add a Scalafix rule or WartRemover config to warn on throw in new files. Prioritize the most critical error paths (payment handling, data mutation) for refactoring first. Track the ratio of exception-based versus typed-error code over time to show progress.

Expect these follow-ups

• How do you convince the team this is worth the investment when they are focused on shipping features?
• What is an acceptable use of throw in a functional Scala codebase?

As asked

A Spark job that has been running for weeks suddenly starts throwing OutOfMemoryError in production. The data volume has not changed significantly. What are the first five things you check?

Sample answer outline

First five checks: (1) Spark UI stage details for which task failed and its memory usage, (2) shuffle spill metrics (spill to disk means not enough memory for shuffle), (3) executor memory configuration versus actual data size per partition, (4) recent code changes that might have introduced a collect or a cartesian join, (5) input data schema changes (new columns or exploded arrays increasing row size). A strong candidate distinguishes between executor OOM and driver OOM and checks different things for each.

Expect these follow-ups

• How do you distinguish between an executor OOM and a driver OOM from the error message?
• What is the fix if the OOM is caused by a single large shuffle partition?

As asked

You see this error: 'ambiguous implicit values: both value x and value y of type Encoder[MyType] match the implicit parameter.' The code was working yesterday. What do you do?

Sample answer outline

First, find what changed: a new import, a new dependency that brought in an implicit, or a new implicit defined in the companion. Use -Xprint:typer or scalacOptions += -Vimplicits to see what the compiler is resolving. Fix options: remove the ambiguous one, make one lower priority via the priority trick, or specify the implicit explicitly with implicitly[Encoder[MyType]]. A strong candidate distinguishes between a real ambiguity (two valid but different instances) and a false one (same instance accessible by two paths).

Expect these follow-ups

• How does making one implicit lower priority using a trait solve the ambiguity?
• How do you prevent this from happening in a library you publish?

As asked

Your team's Scala codebase takes 15 minutes to compile from scratch. New engineers hate it and it is slowing down the development loop. What concrete steps do you take to improve this?

Sample answer outline

Steps: (1) Profile with -Ystatistics or sbt's built-in compilation diagnostics to find the slowest files. (2) Look for implicit search hot spots (a single file triggering thousands of implicit resolutions). (3) Break up monolithic modules into smaller ones so incremental compilation (Zinc) works better. (4) Upgrade to Scala 3 where compile times for implicit search improved. (5) Use build caching (Bazel, Mill, sbt-turbo). Also check for macro expansion overhead from circe or shapeless derivation and consider replacing with explicit codecs.

Expect these follow-ups

• How does Scala 3's givens improve compile time compared to Scala 2 implicits?
• What is the impact of using automatic derivation (shapeless, magnolia) versus manual instances on compile time?

As asked

Your Cats Effect service becomes unresponsive under load. You suspect thread starvation. How do you diagnose this and what are the two most common causes in Cats Effect applications?

Sample answer outline

Diagnose with a thread dump (kill -3 or JFR) to see what threads are blocked on. Common causes: (1) blocking IO (JDBC, file read) run directly on the compute pool instead of wrapped in IO.blocking or evalOn(blockingPool), (2) a CPU-bound computation that does not yield, starving other fibers. Fix: wrap all blocking calls in IO.blocking which runs them on a separate unbounded thread pool, and use IO.cede strategically in long CPU computations. Also enable cats-effect-3's IORuntime diagnostic metrics.

Expect these follow-ups

• What does IO.blocking do differently from IO.delay, and when does it matter?
• How do you configure a dedicated blocking thread pool in a Cats Effect app?

As asked

You need to change the schema of a Kafka topic (adding a required field to a JSON event) without downtime. Producers and consumers are separate Scala services deployed independently. What is your rollout plan?

Sample answer outline

Make the field optional first (add it as Option[T] in the consumer decoder). Deploy consumers that handle both old (field absent) and new (field present) events. Then deploy producers that emit the new field. Once all old messages have been consumed (or consumer groups have caught up past the deployment point), make the field required on the consumer side. Using a schema registry (Confluent or AWS Glue) enforces compatibility. A strong candidate mentions backward versus forward compatibility modes.

Expect these follow-ups

• What is the difference between backward and forward schema compatibility?
• How does Avro with a schema registry handle field addition differently from plain JSON?

As asked

A bug report says users see stale data after an update. Your Scala service caches lookups in a Caffeine cache with a 10-minute TTL. Walk me through how you would confirm the cache is the cause and then fix it correctly.

Sample answer outline

Confirm: add a log entry on cache hit versus cache miss, reproduce the bug, and check logs to see if the stale response was a cache hit. Fix options: reduce TTL (accept higher DB load), invalidate the cache on write (write-through strategy), or use event-driven invalidation (the write service publishes an event, the read service evicts the key). In a distributed deployment with multiple instances, a single-node cache means evicting one node does not fix others, so a distributed cache or event-based invalidation is needed.

Expect these follow-ups

• How would you implement cache invalidation across 10 instances of the same service?
• What are the risks of reducing TTL to zero and using cache-aside without a TTL?

As asked

Tell me about a time you encountered a type-level abstraction in Scala (using implicits, type classes, or higher-kinded types) that was too clever and made the code harder to understand. What did you do?

Sample answer outline

The story should involve a specific over-engineered abstraction: a multi-level type class hierarchy that replaced a simple trait, or a free monad where direct tagless final would have worked. The candidate should explain how they identified the cost (compile times, onboarding difficulty, hard error messages) and then show how they simplified it, ideally with team buy-in. Bonus for mentioning that they kept the simpler version equally type-safe but more readable.

Expect these follow-ups

• How do you draw the line between useful abstraction and accidental complexity?
• What is a type-level feature you actively avoid in production Scala code and why?

As asked

Tell me about a time you helped an engineering team shift from an imperative or object-oriented Scala style to a more functional one. What were the biggest learning curves and how did you address them?

Sample answer outline

Common learning curves: effect systems (why not just Future), error handling with Either versus exceptions, immutability discipline, and reading type errors from implicit resolution. Strong answers show specific interventions: workshops, pair programming sessions, establishing a style guide, and adding Scalafix rules. The result should include a measurable improvement (fewer runtime NullPointerExceptions, faster onboarding of new hires, or improved test coverage).

Expect these follow-ups

• What is the single most impactful functional programming concept for an OO engineer to learn first?
• How do you handle a team member who actively resists the functional style?

As asked

You observe in production that an Akka actor's mailbox is full and messages are being dropped. The actor processes database writes. What are the steps you take to fix this without data loss?

Sample answer outline

First, check why the actor is slow: slow database, too few connections, or a blocking call on the actor dispatcher. Short-term: increase mailbox capacity or switch to a bounded mailbox with backpressure. Long-term: replace the actor with Akka Streams (which has built-in backpressure and prevents mailbox overflow by design) or use a persistent queue (Kafka) so messages are durable before the actor processes them. Never just increase the mailbox indefinitely as this hides the real bottleneck.

Expect these follow-ups

• How do you detect a blocking call on the Akka dispatcher in a production system?
• What is a BoundedMailbox and how does it interact with backpressure?