Every exercise from the twelve modules in one place. The full wording, including the exact
"done when" condition, is in section 6 of each module; this file adds a hint for each one.
Do the easy ones as you read the module, the medium ones the same week, and pick at least two hard
ones per module. The hard ones are what you show in interviews.
Difficulty: E easy (under 30 minutes), M medium (1 to 2 hours), H hard (half a day or more).
Suggested minimum path if time is short: 01.3, 02.4, 02.6, 03.4, 04.3, 04.6, 05.4, 05.6, 06.4,
06.5, 07.1, 07.5, 08.2, 08.5, 09.1, 09.4, 10.4, 11.5, 11.6, 12.6.
pgrep -n sleep; ls -l /proc/PID/fd; after kill PID, wait %1; echo $?. 143 = 128 + 15.
1.4
M
Usernames with /bin/bash shell from /etc/passwd, with awk+sort and with grep+cut.
awk -F: '$7=="/bin/bash"{print $1}'; grep ':/bin/bash$' \| cut -d: -f1. Diff the two outputs with diff <(...) <(...).
1.5
H
p50/p95/p99 from a file of <timestamp> <latency_ms> lines.
Generate data with awk 'BEGIN{srand(); for(i=0;i<10000;i++) print i, int(rand()*100)}'. Sort numerically on field 2, load into an awk array, index at int(p*NR) (clamp to at least 1).
1.6
H
SSH key, GitHub, clone over SSH into ~/projects, core.autocrlf input.
ssh-keygen -t ed25519, paste ~/.ssh/id_ed25519.pub into GitHub settings; ssh -T git@github.com. Clone into the Linux filesystem, not /mnt/c.
g++ -std=c++17 -E hello.cpp \| wc -l. The difference is every header <iostream> pulls in transitively: templates, locales, strings.
2.2
E
Build m02_static_lib, find the U that becomes T with nm -C.
Run build.sh; compare nm -C main.o \| grep add and nm -C prog_static \| grep add.
2.3
M
Make prog_norpath run three ways; rank them for production.
LD_LIBRARY_PATH=.; sudo cp libmathlib.so /usr/local/lib && sudo ldconfig; relink with -Wl,-rpath,'$ORIGIN' (single quotes so the shell does not eat $ORIGIN).
2.4
M
Reproduce the three undefined-reference cases in m02_undefined_ref; diagnose each from nm -C.
Missing object: symbol absent everywhere. C vs C++ linkage: nm shows add in one file and add(int, int) in the other (use nm without -C to see _Z3addii). Signature mismatch: two different mangled names.
2.5
H
libplugin.so with one exported extern "C" function, loaded via dlopen/dlsym; -fvisibility=hidden.
Mark the function __attribute__((visibility("default"))); link the host with -ldl (not needed on glibc 2.34+, harmless). nm -D --defined-only libplugin.so.
2.6
H
silent_odr links and misbehaves; -flto -Wodr reports it; fix it.
The linker only sees symbol names, not class layouts, so two different struct Config definitions with the same name both "link". The fix is one shared header.
Without ASan the overrun lands in malloc padding or another chunk, so nothing visible happens. ASan's report has three stacks: the bad access, the allocation, and (for use-after-free) the free.
3.5
H
Signed overflow in a loop condition: terminates at -O0, infinite at -O2; then UBSan.
for (int i = 1; i > 0; i *= 2) or i += big. The optimiser assumes signed overflow never happens, so i > 0 is "always true". timeout 2 ./prog; echo $? gives 124.
3.6
H
-static-libstdc++ -static-libgcc, then fully -static; run in a FROM scratch image.
Dockerfile: FROM scratch, COPY hello /hello, ENTRYPOINT ["/hello"]. ldd on the fully static binary says "not a dynamic executable".
target_include_directories(... PUBLIC ...), Threads::Threads, CMAKE_CXX_STANDARD + CXX_EXTENSIONS OFF.
4.3
M
Flip Threads::Threads and the include dir from PUBLIC to PRIVATE; explain what breaks.
glibc 2.34 moved pthread into libc, so the link may still succeed; the include change breaks main.cpp at compile time. Rule: PUBLIC if it appears in your headers, PRIVATE if only in your .cpp.
4.4
M
A clang preset; fix what Clang warns about.
The preset exists in the shipped CMakePresets.json; build it with cmake --preset clang && cmake --build --preset clang. Clang often adds -Wshorten-64-to-32-style findings.
4.5
H
Second library tcpkv_net depending on tcpkv_core; tcpkv depends on tcpkv_net.
install(EXPORT), install to /tmp/tcpkv-install, consume with find_package(tcpkv CONFIG).
You need a tcpkvConfig.cmake that include()s the exported targets file; configure_package_config_file + write_basic_package_version_file from CMakePackageConfigHelpers.
SET with : and - in the key passes; a tab in the key fails.
Add to the parameterised MalformedLines table for the tab case; check ctest -N \| grep -i tab.
5.2
E
--gtest_shuffle --gtest_repeat=50; a shared static makes order matter.
Put static int counter in the test file, increment in test A, assert in test B. Shuffle exposes it within a few repeats; --gtest_random_seed=N reproduces.
5.3
M
Store::require_key aborts on empty key; EXPECT_DEATH.
Death tests fork; name the suite *DeathTest so they run first. Matcher is a regex on stderr.
5.4
M
Remove the lock in Store::get; read the TSan report.
cmake --preset tsan && cmake --build --preset tsan && ctest --preset tsan -R Concurrency. The report shows "Write of size ... by thread T2" and "Previous read ... by thread T1" with both stacks.
5.5
H
Extract a LineReader with tests for split lines, CRLF and overflow.
Keep the buffer and a start offset; feed appends, scans for \n, returns complete lines, erases the consumed prefix. Test overflow at exactly kMaxLineLength and one byte more.
5.6
H
ctest integration test with FIXTURES_SETUP/FIXTURES_CLEANUP and the Python client.
Setup test: a script that starts tcpkv on 7123 in the background and writes its PID to a file; cleanup test: kill $(cat pidfile); main test: tcpkv_client.py --selftest --port 7123 --wait 5. Label all three integration.
git cat-file -p HEAD, copy the tree hash, git cat-file -p <tree>, copy a blob hash. git cat-file -t prints the type.
6.2
E
Stage half a file's hunks with git add -p; two commits on one file.
In add -p, s splits a hunk, e edits it by hand.
6.3
M
--fixup plus rebase -i --autosquash non-interactively.
git commit --fixup=<sha-of-first>; GIT_SEQUENCE_EDITOR=true git rebase -i --autosquash <base>.
6.4
M
Conflict with zdiff3, resolve, then undo the whole rebase via reflog.
git config merge.conflictstyle zdiff3. The middle region is the common ancestor. Undo: git reflog, find the entry before rebase (start), git reset --hard HEAD@{n}.
The finished script is in code/tcpkv/scripts/bisect_test.sh; copy it outside the repo first so old commits can run it. Plant the bug with git commit in a loop of trivial changes before and after.
Add -fno-omit-frame-pointer to the release preset; perf record -g -p $(pgrep -n tcpkv) -- sleep 15 while scripts/load_test.sh 7000 8 20000 runs; Brendan Gregg's FlameGraph scripts render the SVG.
7.4
M
callgrind on ParseCommand.*; allocation share of the parser.
valgrind --tool=callgrind ./build/debug/tests/tcpkv_tests --gtest_filter='ParseCommand.*'; callgrind_annotate --inclusive=yes callgrind.out.*; compare parse_command inclusive with operator new under it.
7.5
H
Lock-order deadlock: gdb thread apply all bt, /proc/PID/task/*/wchan, TSan report.
gdb -batch -x scripts/dump_stacks.gdb -p PID dumps all stacks non-interactively. In each blocked thread, p mutex shows __owner = the LWP holding it.
7.6
H
tcpdump -w a tcpkv session, read in Wireshark and with tcpdump -A.
sudo tcpdump -i lo -w kv.pcap port 7000; copy to /mnt/c/... for Windows Wireshark; tcpdump -r kv.pcap -A for the text view. Flags: [S], [S.], [P.], [F.].
docker build and docker history; name the three largest layers.
docker history --no-trunc --format '{{.Size}}\t{{.CreatedBy}}' tcpkv:dev. The apt layers dominate.
8.2
E
docker stop shows exit 0; without sigaction, exit 137 after 10 s.
docker inspect -f '{{.State.ExitCode}}' tcpkv. With tini as PID 1, tini forwards SIGTERM and the unhandled default action kills tcpkv at once (143). You see 137 after 10 s when tcpkv itself is PID 1 (--entrypoint /usr/local/bin/tcpkv): the kernel does not apply default "terminate" actions to a namespace's PID 1, so Docker times out and sends SIGKILL. Try both.
8.3
M
Runtime stage on debian:12-slim and distroless; strace via a sidecar.
docker run --rm -it --pid=container:tcpkv --cap-add SYS_PTRACE ubuntu:24.04 then apt install strace; strace -p 1.
8.4
M
HEALTHCHECK goes (healthy), then (unhealthy) after docker kill -s STOP.
The shipped Dockerfile already has the healthcheck; docker ps shows the state; docker inspect -f '{{json .State.Health}}' shows the probe log.
8.5
H
Compose stack; psql and redis-cli from the host; OOM-kill tcpkv with large values.
psql -h 127.0.0.1 -U tcpkv; set the compose memory limit to 64M; scripts/load_test.sh 7000 4 50000 65000; docker inspect -f '{{.State.OOMKilled}}' tcpkv.
8.6
H
Fully static tcpkv in a FROM scratch image.
-static with glibc warns about getaddrinfo (it needs NSS modules at run time); tcpkv only binds to an IP, so it works. Remove the nc-based healthcheck: scratch has no shell.
A PR with an unused variable is blocked by -Werror.
gh pr create, then gh run view --log-failed and search for -Werror=unused-variable.
9.3
M
format and tidy jobs with .clang-format and .clang-tidy.
clang-format --dry-run --Werror $(git ls-files '*.cpp' '*.hpp'); clang-tidy -p build/ci src/*.cpp. Configure first so compile_commands.json exists.
9.4
M
Smoke-test the built image in the docker job.
Build with load: true in build-push-action, then docker run -d -p 7000:7000, loop on printf 'PING\n' \| nc -N localhost 7000 \| grep -q PONG with a timeout, docker rm -f.
9.5
H
Nightly workflow (full Valgrind, -O3 -flto) and a tag-triggered release with two assets.
Port the workflow to Azure DevOps; JUnit results visible in the UI.
The file is code/tcpkv/azure-pipelines.yml; a free Azure DevOps organisation includes Microsoft-hosted minutes for public projects (check current limits). Gate the Docker stage on a variable if you have no registry.
sudo apt install shellcheck. The PS4 variable defines the -x prefix: file, line, function.
10.2
E
A CRLF script fails with bash\r; fix with dos2unix; prevent with .gitattributes.
file hello.sh reports "with CRLF line terminators"; cat -A shows ^M$. Rule: *.sh text eol=lf.
10.3
M
Demonstrate each set -e hole (if, local x=$(false), ((x++)), function under \|\|).
Put an echo "still running" after each; the ones that print are holes. Fix: separate declaration from assignment, x=$((x+1)), explicit checks inside functions.
10.4
M
Ctrl-C during run_tests.sh leaves no server behind.
The INT trap calls exit 130, which runs the EXIT trap (stop_server) once. Check with pgrep -a tcpkv.
One TcpkvClient per thread; random.choices with weights for the op type; per-op latency lists merged at the end. It is slower than load_test.sh because every request waits for its reply (no pipelining) and Python adds overhead per call.
10.6
H
ctest integration fixture using the Python client.
Same as 5.6. Write a small scripts/start_server.sh and register it with add_test(NAME start_server COMMAND ${CMAKE_SOURCE_DIR}/scripts/start_server.sh $<TARGET_FILE:tcpkv> 7123) so ctest passes the real binary path via a generator expression.
Each m11_syscalls demo under strace -f; name the syscalls behind open, fork, waitpid.
openat, clone (or clone3), wait4. clock_gettime runs in the vDSO without entering the kernel.
11.2
E
ulimit -n 64, start tcpkv, open 100 connections; see EMFILE.
Python: socks = [socket.create_connection(("127.0.0.1", 7000)) for _ in range(100)] in a try block. cat /proc/$(pgrep -n tcpkv)/limits.
11.3
M
Measure the cost of fsync; write the atomic-replace version.
hyperfine './w_nosync' './w_sync'. Directory fsync: int d = open(".", O_RDONLY \| O_DIRECTORY); fsync(d);. WSL's virtual disk makes absolute numbers unrepresentative; compare ratios.
11.4
M
mini_shell.c with one pipe and one > redirect.
Fork twice for a pipeline; in each child dup2 the right pipe end then close both ends; the parent closes both ends and waits for both children. Forgetting a close makes wc wait forever.
11.5
M
tcpkv as a systemd service; crash it; exceed StartLimitBurst.
sudo kill -SEGV $(systemctl show -p MainPID --value tcpkv). After the burst: systemctl status says start-limit-hit; sudo systemctl reset-failed tcpkv clears it.
11.6
H
epoll-based tcpkv server in one thread with output buffers and signalfd.
Start from m11_epoll_preview.c (rename to .cpp, or link a C file with the C++ core through a small extern "C" layer). Per connection: std::string in, out; enable EPOLLOUT only while out is non-empty. See 08_Networking_Distributed/06_nonblocking_io_epoll_event_loops.md.
cl build plus dumpbin /headers /dependents /pdbpath; the same with g++, readelf, ldd.
Use the "Developer PowerShell for VS 2022" so cl and dumpbin are on PATH.
12.2
M
tcpkv core and tests under MSVC /W4 /WX; fix warnings portably.
Configure with -G "Visual Studio 17 2022" -A x64 and build only tcpkv_tests. Prefer static_cast and correct types (std::size_t) over #pragma warning(disable).
12.3
M
Reproduce and fix LNK2038 (/MT library, /MD exe).
set_property(TARGET lib PROPERTY MSVC_RUNTIME_LIBRARY MultiThreaded) on one target only reproduces it; set CMAKE_MSVC_RUNTIME_LIBRARY once at the top to fix it.
12.4
M
vcpkg.json with fmt, built on Windows and WSL.
Pass -DCMAKE_TOOLCHAIN_FILE=$VCPKG_ROOT/scripts/buildsystems/vcpkg.cmake before the first configure (it cannot be added to an existing cache). Triplets: x64-windows and x64-linux.
12.5
H
Port main.cpp to Winsock with the socket_t shim; same Python self-test passes on both.
Replace poll with WSAPoll, accept4 with accept, MSG_NOSIGNAL with 0, localtime_r with localtime_s (argument order differs). SetConsoleCtrlHandler runs on its own thread: set an std::atomic<bool>.
12.6
H
Crash dump on both platforms: ProcDump + WinDbg, coredumpctl + GDB.
WinDbg: .symfix; .sympath+ <dir with your pdb>; .reload; !analyze -v; .ecxr; kp. Linux: coredumpctl list, coredumpctl gdb tcpkv, bt full.