Low-Latency C++ Developer Interview Questions & Answers
40 questions for Low-Latency C++ Developer interviews — 17 medium, 23 hard.
About the Low-Latency C++ Developer role
Low-latency developers at trading firms and investment banks build market data handlers, order gateways and strategy engines where microseconds matter. They use lock-free queues, atomics, cache-aware data structures, kernel bypass networking and careful memory management. Interviews are demanding and cover advanced C++, concurrency, memory ordering, templates and system design. Strong problem solvers with deep C++ and Linux knowledge earn some of the highest packages in India.
Lock-free programming Memory ordering Linux internals Cache optimisation Network programming
Solved for this roleStart practicing
3 QUESTIONS FREESign up to unlock the full question bank and track your progress.Unlock every question and answer, then continue practising company-wise and package-wise.
✓ 3 questions free • Unlock the rest in seconds
01
How does std::condition_variable work and what are spurious wakeups?
A condition variable lets a thread sleep until another thread signals that some state changed. The waiter holds a unique_lock and calls cv.wait(lock, predicate), which releases the mutex while sleeping and reacquires it on wake. The notifier changes the state under the mutex and calls notify_one or notify_all. Threads can wake without notification, called spurious wakeups, so always wait with a predicate that rechecks the condition.
02
Design a producer-consumer pipeline where one thread reads log lines and several workers parse them.
I would use a bounded thread-safe queue built with a mutex and two condition variables, one for not empty and one for not full, so a fast reader cannot exhaust memory. The reader pushes lines and workers pop and parse them. For shutdown, the reader pushes a sentinel per worker or sets a closed flag and notifies all. Batching lines per push reduces locking overhead, and results go to a second queue.
03
Compare std::mutex, lock_guard, unique_lock and scoped_lock.
std::mutex provides lock and unlock, but calling them manually is error-prone. lock_guard is a simple RAII wrapper that locks in the constructor and unlocks in the destructor. unique_lock is more flexible: it can defer locking, unlock early, be moved and is required by condition_variable. scoped_lock, from C++17, locks one or more mutexes at once using a deadlock-avoidance algorithm.
04
What is a custom deleter for a smart pointer and when do you need one?
A custom deleter tells the smart pointer how to release a resource that was not created with plain new. For example std::unique_ptr<FILE, decltype(&fclose)> f(fopen(path, mode), &fclose); closes the file automatically. Deleters are used for C library handles, sockets, OS handles and objects from custom pools. For unique_ptr the deleter is part of the type, while shared_ptr stores it in the control block.
05
The release binary has grown very large and analysis shows many template instantiations. What can you do?
Each distinct set of template arguments generates separate code, so heavy generic code bloats binaries. I would find the biggest symbols with tools like bloaty or nm, move type-independent logic into a non-template base or helper, use type erasure for rarely hot paths, limit instantiations with extern template, and enable link-time optimisation plus identical code folding to merge duplicate functions.
06
A teammate complains that a template function produces hundreds of lines of compiler errors when misused. How do you improve this?
Constrain the template so misuse fails early with a clear message. With C++20 I would add a concept or requires clause, for example template <std::totally_ordered T>, so the error names the unmet requirement at the call site. On older standards, a static_assert with a readable message at the top of the function body helps. Good naming and a comment on expected type properties also help.
07
What are type traits and how does if constexpr use them?
Type traits in <type_traits> answer questions about types at compile time, such as std::is_integral_v<T>, std::is_pointer_v<T> or std::is_same_v<T, U>, and transform types, such as std::remove_reference_t. if constexpr picks a branch at compile time and discards the other, so code like if constexpr (std::is_integral_v<T>) can call integer-only operations safely inside a generic function.
08
You need a container that holds ints, strings and doubles together. Which approach would you choose?
If the set of types is fixed and known, std::vector<std::variant<int, std::string, double>> is type-safe and needs no heap per element; std::visit handles each type. If types are open-ended, std::any works but needs any_cast and loses compile-time checks. If the items share behaviour, a base class interface with unique_ptr is cleanest. I usually prefer variant for closed sets.
Placement new constructs an object in memory you already own, written new (buffer) T(args). It does not allocate, so you must call the destructor yourself and release the buffer separately. A memory pool pre-allocates large blocks and hands out fixed-size slots, avoiding slow general-purpose allocation and fragmentation. Game engines, low-latency trading systems and std::pmr allocators use these for predictable performance.
10
Tell me about a time you wrote generic, reusable code that saved effort for your team.
Situation: three services each had their own copy of retry logic for network calls. Task: remove duplication without slowing them down. Action: I wrote a small header-only template, retry with a callable and a policy for attempts and backoff, added unit tests and documented examples. Result: all three services adopted it, around 600 lines of duplicate code were deleted and a bug fix now happens in one place.
11
Write a micro-benchmark with Google Benchmark comparing two approaches.
I would measure first with -ftime-trace or build analyzers to find the costly headers. Then reduce includes with forward declarations, move template-heavy implementation details behind non-template interfaces, use extern template with explicit instantiation for common types, enable precompiled headers and a compiler cache such as ccache, and evaluate C++20 modules where the toolchain supports them well.
13
Many threads read a shared configuration object constantly, but it changes only once an hour. How do you synchronise it efficiently?
A plain mutex would serialise all readers. Option one is std::shared_mutex: readers take shared_lock and run in parallel, the updater takes unique_lock. Option two, often faster, is to keep the config in a std::shared_ptr<const Config> and swap in a new immutable object atomically with std::atomic<std::shared_ptr> in C++20; readers grab a copy of the pointer and never block the writer.
Situation: a market data service dropped a few updates per day only under heavy load. Task: find the cause without disrupting trading hours. Action: I wrote a stress harness replaying recorded traffic, ran it under ThreadSanitizer and found two threads updating a shared index without synchronisation. I redesigned it to a single-writer model with an SPSC queue. Result: zero dropped updates over a month and latency improved slightly because locks were removed.
16
Write a GoogleTest unit test for a function, including an edge case.
It depends on ownership. If the function only uses the object, take T& or T* and let the caller manage lifetime. Take shared_ptr by value only when the function stores or shares ownership, and then move it into place. Passing shared_ptr by value everywhere adds atomic reference count updates that slow down hot paths and hide who really owns the object.
Substitution Failure Is Not An Error: when the compiler substitutes template arguments and the resulting declaration is invalid, that candidate is silently removed from overload resolution instead of causing an error. It was traditionally used with std::enable_if to enable functions only for certain types, such as integral types. In C++20, concepts and requires clauses replace most SFINAE tricks with clearer syntax and error messages.
Templates give static polymorphism: resolved at compile time, inlinable and fast, but types must be known at compile time and code size grows. Virtual functions give dynamic polymorphism: one compiled body, types can be chosen at run time, plugins are possible, at the cost of an indirect call. I use templates for performance-critical generic algorithms and virtual interfaces for runtime-configurable components.
22
A low-latency service shows occasional latency spikes of several milliseconds. How would you investigate?
I would first capture latency percentiles and correlate spikes with events. Common causes are heap allocation in the hot path, page faults, lock contention, logging that blocks on I/O, CPU frequency changes and the OS scheduling other work on the same core. Tools include perf, off-CPU flame graphs and tracing. Typical fixes are preallocating memory, pinning threads to isolated cores, async logging and avoiding syscalls in hot paths.
23
A production service's memory usage grows steadily over several days. How do you find the cause?
First confirm it is a real leak and not a cache or allocator fragmentation by tracking RSS and allocator statistics. Then reproduce in staging with a heap profiler such as heaptrack or jemalloc profiling and compare snapshots to see which call stacks keep growing. LeakSanitizer catches lost blocks. Common culprits are unbounded maps, shared_ptr cycles and listeners never removed. After the fix, add a long soak test.
24
A low-latency system needs a queue between two threads with no locks. How would you approach it?
For one producer and one consumer, a single-producer single-consumer ring buffer is the standard answer: a fixed-size array with atomic head and tail indices, using release on publish and acquire on read, and padding so the indices sit on different cache lines. It avoids allocation and syscalls. For multiple producers I would use a proven library rather than writing my own, and verify with stress tests and TSan.
25
Compute the sum of a large vector in parallel using std::async.
#include <future>
#include <numeric>
#include <vector>
long long parallelSum(const std::vector<int>& v, int parts = 4) {
std::vector<std::future<long long>> futs;
std::size_t chunk = v.size() / parts;
for (int i = 0; i < parts; ++i) {
auto b = v.begin() + i * chunk;
auto e = (i == parts - 1) ? v.end() : b + chunk;
futs.push_back(std::async(std::launch::async, [b, e] {
return std::accumulate(b, e, 0LL);
}));
}
long long total = 0;
for (auto& f : futs) total += f.get();
return total;
}
// C++17 alternative: std::reduce(std::execution::par, v.begin(), v.end(), 0LL)
26
Tell me about a code review where you pointed out an important issue in a colleague's code.
Situation: a colleague's change stored pointers to vector elements in a map and then appended to the vector. Task: flag the risk without discouraging them. Action: I commented with a short example showing reallocation invalidating the pointers, suggested storing indices or using std::deque, and offered to pair on the fix. Result: the bug was fixed before release, and we added a checklist item on iterator invalidation for reviews.
The program depends on a shared library that is missing or in a different version on the server, or the loader cannot find it. I would run ldd on the binary to see unresolved libraries, then install the right package, set the rpath at link time, or ship the library alongside. A newer glibc or libstdc++ on the build machine is a common cause; build inside a container matching production, or link statically where allowed.
29
Which profiling tools do you use for C++ and what does each show?
On Linux, perf samples the CPU and shows hot functions and call graphs, often visualised as flame graphs. Valgrind callgrind counts instructions precisely but runs slowly, and massif or heaptrack profile memory. gprof is older and needs instrumented builds. Intel VTune and AMD uProf show cache misses and pipeline stalls. On Windows, the Visual Studio profiler covers CPU and memory.
30
What is template argument deduction and class template argument deduction?
For function templates the compiler deduces T from the call, so maxOf(3, 7) deduces int; mixing int and double fails because T would be ambiguous. Class template argument deduction, added in C++17, lets you write std::vector v{1, 2, 3}; or std::pair p(1, 2.5); without spelling the template arguments. Deduction guides can be written to tell the compiler how to deduce types from constructors.
31
What is the difference between full and partial template specialization?
Full specialization provides a separate implementation for one exact set of template arguments, such as template <> class Hash<std::string>. Partial specialization, allowed only for class and variable templates, covers a family of arguments, such as template <typename T> class Hash<T*> for all pointer types. Function templates cannot be partially specialised; overloading is used instead.
A variadic template accepts any number of template arguments using a parameter pack, written template <typename... Args>. Before C++17 packs were processed recursively. Fold expressions from C++17 apply an operator across the whole pack in one line, for example return (args + ...); sums all arguments. They power std::make_unique, std::tuple, std::format and logging functions that take arbitrary arguments.
33
Describe a time you improved code quality by changing how memory was managed.
Situation: our module used raw new and delete everywhere and QA reported frequent leaks. Task: make it safe without a full rewrite. Action: I proposed a step-by-step migration to unique_ptr and std::vector, wrote a one-page ownership guideline and enabled LeakSanitizer in CI so new leaks failed the build. Result: leak reports dropped to zero within two sprints and reviews became faster because ownership was obvious.
34
What is std::atomic and what do memory orderings mean?
std::atomic<T> makes reads, writes and read-modify-write operations such as fetch_add indivisible and race-free without a mutex. Memory order controls how other memory operations are ordered around it. seq_cst, the default, gives one global order and is easiest to reason about. acquire and release pair up to publish data from one thread to another. relaxed only guarantees atomicity, suitable for simple counters. Weaker orders are faster but harder to get right.
35
A C++ server crashes with a segmentation fault in production. How do you debug it?
Make sure core dumps are enabled with ulimit -c unlimited or systemd-coredump, and keep debug symbols for release builds. Load the core with gdb binary core, run bt to see the stack trace, then inspect frames and variables. If the stack looks corrupted, reproduce in staging with AddressSanitizer. I would also check logs around the crash time and recent deployments, then add a regression test with the fix.
First switch the compiler standard flag and fix build errors, with CI running the full test suite. Then add clang-tidy modernize checks to convert NULL to nullptr, add override and use auto where it helps, applied module by module in small reviewed batches. Next replace raw owning pointers with unique_ptr in new and touched code. I avoid a big-bang rewrite, track progress with metrics and train the team as we go.
38
Write a variadic function that sums any number of arguments and one that prints them.
They are compiler instrumentation tools in GCC and Clang. ASan, -fsanitize=address, detects out-of-bounds access, use-after-free, double free and leaks with about 2x slowdown. UBSan, -fsanitize=undefined, catches signed overflow, invalid shifts and null dereferences. TSan, -fsanitize=thread, detects data races but cannot be combined with ASan. Running tests under sanitizers in CI catches bugs long before production.
40
You parallelised a loop across eight threads but it became slower than the single-threaded version. What could be wrong?
Likely causes are lock contention when all threads share one mutex, false sharing when per-thread counters sit on the same 64-byte cache line, too little work per task so thread creation dominates, or memory bandwidth limits. I would profile with perf, pad per-thread data with alignas(64) or std::hardware_destructive_interference_size, use thread-local accumulation and a thread pool, and make chunks larger.
Frequently asked questions
What does a Low-Latency C++ Developer do?
Low-latency developers at trading firms and investment banks build market data handlers, order gateways and strategy engines where microseconds matter. They use lock-free queues, atomics, cache-aware data structures, kernel bypass networking and careful memory management. Interviews are demanding and cover advanced C++, concurrency, memory ordering, templates and system design. Strong problem solvers with deep C++ and Linux knowledge earn some of the highest packages in India.
Which skills are tested in a Low-Latency C++ Developer interview?
Expect questions on Lock-free programming, Memory ordering, Linux internals, Cache optimisation, Network programming. Most rounds also ask you to explain a project and solve one practical problem live.
How do I prepare for a Low-Latency C++ Developer interview?
Revise Multithreading and Concurrency, Performance, Debugging and Build Tools, Templates and Generic Programming, Pointers and Memory Management first using the topic sheets, practise the questions on this page aloud, then do one full mock interview to practise timing.