ClickHouse/docker/test/stress/run.sh
Azat Khuzhin ebed6193b4 tests/stress: add core dumps into artifacts
gcore is used here because:

- default kernel.core_pattern is "|/usr/share/apport/apport %p %s %c %d %P %E" [1]
  and so to make it work you need to install apport into container and
  configure it propertly -- to complex

  [1]: https://s3.amazonaws.com/clickhouse-test-reports/33389/204e459d259570e6bc7fe2903f7e516094a916ca/stress_test__address__actions_/runlog.log

- kernel.core_pattern cannot be changed, since proc is mounted in
  read-only mode in non --privileged containers [2]

  [2]: https://s3.amazonaws.com/clickhouse-test-reports/33389/dab3afbd94558f9654bf0aa1e06e06e2962f3bb0/stress_test__address__actions_/runlog.log

v2: change kernel.core_pattern
v3: use gcore, since you are not allowed to change kernel.core_pattern
    in container (only under --privileged)
2022-01-04 14:03:40 +03:00

274 lines
12 KiB
Bash
Executable File

#!/bin/bash
# shellcheck disable=SC2094
# shellcheck disable=SC2086
# shellcheck disable=SC2024
set -x
# Thread Fuzzer allows to check more permutations of possible thread scheduling
# and find more potential issues.
export THREAD_FUZZER_CPU_TIME_PERIOD_US=1000
export THREAD_FUZZER_SLEEP_PROBABILITY=0.1
export THREAD_FUZZER_SLEEP_TIME_US=100000
export THREAD_FUZZER_pthread_mutex_lock_BEFORE_MIGRATE_PROBABILITY=1
export THREAD_FUZZER_pthread_mutex_lock_AFTER_MIGRATE_PROBABILITY=1
export THREAD_FUZZER_pthread_mutex_unlock_BEFORE_MIGRATE_PROBABILITY=1
export THREAD_FUZZER_pthread_mutex_unlock_AFTER_MIGRATE_PROBABILITY=1
export THREAD_FUZZER_pthread_mutex_lock_BEFORE_SLEEP_PROBABILITY=0.001
export THREAD_FUZZER_pthread_mutex_lock_AFTER_SLEEP_PROBABILITY=0.001
export THREAD_FUZZER_pthread_mutex_unlock_BEFORE_SLEEP_PROBABILITY=0.001
export THREAD_FUZZER_pthread_mutex_unlock_AFTER_SLEEP_PROBABILITY=0.001
export THREAD_FUZZER_pthread_mutex_lock_BEFORE_SLEEP_TIME_US=10000
export THREAD_FUZZER_pthread_mutex_lock_AFTER_SLEEP_TIME_US=10000
export THREAD_FUZZER_pthread_mutex_unlock_BEFORE_SLEEP_TIME_US=10000
export THREAD_FUZZER_pthread_mutex_unlock_AFTER_SLEEP_TIME_US=10000
dpkg -i package_folder/clickhouse-common-static_*.deb
dpkg -i package_folder/clickhouse-common-static-dbg_*.deb
dpkg -i package_folder/clickhouse-server_*.deb
dpkg -i package_folder/clickhouse-client_*.deb
dpkg -i package_folder/clickhouse-test_*.deb
function configure()
{
# install test configs
/usr/share/clickhouse-test/config/install.sh
# avoid too slow startup
sudo cat /etc/clickhouse-server/config.d/keeper_port.xml | sed "s|<snapshot_distance>100000</snapshot_distance>|<snapshot_distance>10000</snapshot_distance>|" > /etc/clickhouse-server/config.d/keeper_port.xml.tmp
sudo mv /etc/clickhouse-server/config.d/keeper_port.xml.tmp /etc/clickhouse-server/config.d/keeper_port.xml
sudo chown clickhouse /etc/clickhouse-server/config.d/keeper_port.xml
sudo chgrp clickhouse /etc/clickhouse-server/config.d/keeper_port.xml
# for clickhouse-server (via service)
echo "ASAN_OPTIONS='malloc_context_size=10 verbosity=1 allocator_release_to_os_interval_ms=10000'" >> /etc/environment
# for clickhouse-client
export ASAN_OPTIONS='malloc_context_size=10 allocator_release_to_os_interval_ms=10000'
# since we run clickhouse from root
sudo chown root: /var/lib/clickhouse
# Set more frequent update period of asynchronous metrics to more frequently update information about real memory usage (less chance of OOM).
echo "<clickhouse><asynchronous_metrics_update_period_s>1</asynchronous_metrics_update_period_s></clickhouse>" \
> /etc/clickhouse-server/config.d/asynchronous_metrics_update_period_s.xml
local total_mem
total_mem=$(awk '/MemTotal/ { print $(NF-1) }' /proc/meminfo) # KiB
total_mem=$(( total_mem*1024 )) # bytes
# Set maximum memory usage as half of total memory (less chance of OOM).
#
# But not via max_server_memory_usage but via max_memory_usage_for_user,
# so that we can override this setting and execute service queries, like:
# - hung check
# - show/drop database
# - ...
#
# So max_memory_usage_for_user will be a soft limit, and
# max_server_memory_usage will be hard limit, and queries that should be
# executed regardless memory limits will use max_memory_usage_for_user=0,
# instead of relying on max_untracked_memory
local max_server_mem
max_server_mem=$((total_mem*75/100)) # 75%
echo "Setting max_server_memory_usage=$max_server_mem"
cat > /etc/clickhouse-server/config.d/max_server_memory_usage.xml <<EOL
<clickhouse>
<max_server_memory_usage>${max_server_mem}</max_server_memory_usage>
</clickhouse>
EOL
local max_users_mem
max_users_mem=$((total_mem*50/100)) # 50%
echo "Setting max_memory_usage_for_user=$max_users_mem"
cat > /etc/clickhouse-server/users.d/max_memory_usage_for_user.xml <<EOL
<clickhouse>
<profiles>
<default>
<max_memory_usage_for_user>${max_users_mem}</max_memory_usage_for_user>
</default>
</profiles>
</clickhouse>
EOL
}
function stop()
{
clickhouse stop
}
function start()
{
# Rename existing log file - it will be more convenient to read separate files for separate server runs.
if [ -f '/var/log/clickhouse-server/clickhouse-server.log' ]
then
log_file_counter=1
while [ -f "/var/log/clickhouse-server/clickhouse-server.log.${log_file_counter}" ]
do
log_file_counter=$((log_file_counter + 1))
done
mv '/var/log/clickhouse-server/clickhouse-server.log' "/var/log/clickhouse-server/clickhouse-server.log.${log_file_counter}"
fi
counter=0
until clickhouse-client --query "SELECT 1"
do
if [ "$counter" -gt 240 ]
then
echo "Cannot start clickhouse-server"
cat /var/log/clickhouse-server/stdout.log
tail -n1000 /var/log/clickhouse-server/stderr.log
tail -n100000 /var/log/clickhouse-server/clickhouse-server.log | grep -F -v -e '<Warning> RaftInstance:' -e '<Information> RaftInstance' | tail -n1000
break
fi
# use root to match with current uid
clickhouse start --user root >/var/log/clickhouse-server/stdout.log 2>>/var/log/clickhouse-server/stderr.log
sleep 0.5
counter=$((counter + 1))
done
# Set follow-fork-mode to parent, because we attach to clickhouse-server, not to watchdog
# and clickhouse-server can do fork-exec, for example, to run some bridge.
# Do not set nostop noprint for all signals, because some it may cause gdb to hang,
# explicitly ignore non-fatal signals that are used by server.
# Number of SIGRTMIN can be determined only in runtime.
RTMIN=$(kill -l SIGRTMIN)
echo "
set follow-fork-mode parent
handle SIGHUP nostop noprint pass
handle SIGINT nostop noprint pass
handle SIGQUIT nostop noprint pass
handle SIGPIPE nostop noprint pass
handle SIGTERM nostop noprint pass
handle SIGUSR1 nostop noprint pass
handle SIGUSR2 nostop noprint pass
handle SIG$RTMIN nostop noprint pass
info signals
continue
gcore
backtrace full
info locals
info registers
disassemble /s
up
info locals
disassemble /s
up
info locals
disassemble /s
p \"done\"
detach
quit
" > script.gdb
# FIXME Hung check may work incorrectly because of attached gdb
# 1. False positives are possible
# 2. We cannot attach another gdb to get stacktraces if some queries hung
gdb -batch -command script.gdb -p "$(cat /var/run/clickhouse-server/clickhouse-server.pid)" | ts '%Y-%m-%d %H:%M:%S' >> /test_output/gdb.log &
sleep 5
# gdb will send SIGSTOP, spend some time loading debug info and then send SIGCONT, wait for it (up to send_timeout, 300s)
time clickhouse-client --query "SELECT 'Connected to clickhouse-server after attaching gdb'" ||:
}
configure
start
# shellcheck disable=SC2086 # No quotes because I want to split it into words.
/s3downloader --url-prefix "$S3_URL" --dataset-names $DATASETS
chmod 777 -R /var/lib/clickhouse
clickhouse-client --query "ATTACH DATABASE IF NOT EXISTS datasets ENGINE = Ordinary"
clickhouse-client --query "CREATE DATABASE IF NOT EXISTS test"
stop
start
clickhouse-client --query "SHOW TABLES FROM datasets"
clickhouse-client --query "SHOW TABLES FROM test"
clickhouse-client --query "RENAME TABLE datasets.hits_v1 TO test.hits"
clickhouse-client --query "RENAME TABLE datasets.visits_v1 TO test.visits"
clickhouse-client --query "SHOW TABLES FROM test"
./stress --hung-check --drop-databases --output-folder test_output --skip-func-tests "$SKIP_TESTS_OPTION" \
&& echo -e 'Test script exit code\tOK' >> /test_output/test_results.tsv \
|| echo -e 'Test script failed\tFAIL' >> /test_output/test_results.tsv
stop
start
clickhouse-client --query "SELECT 'Server successfully started', 'OK'" >> /test_output/test_results.tsv \
|| echo -e 'Server failed to start\tFAIL' >> /test_output/test_results.tsv
[ -f /var/log/clickhouse-server/clickhouse-server.log ] || echo -e "Server log does not exist\tFAIL"
[ -f /var/log/clickhouse-server/stderr.log ] || echo -e "Stderr log does not exist\tFAIL"
# Print Fatal log messages to stdout
zgrep -Fa " <Fatal> " /var/log/clickhouse-server/clickhouse-server.log*
# Grep logs for sanitizer asserts, crashes and other critical errors
# Sanitizer asserts
grep -Fa "==================" /var/log/clickhouse-server/stderr.log | grep -v "in query:" >> /test_output/tmp
grep -Fa "WARNING" /var/log/clickhouse-server/stderr.log >> /test_output/tmp
zgrep -Fav "ASan doesn't fully support makecontext/swapcontext functions" /test_output/tmp > /dev/null \
&& echo -e 'Sanitizer assert (in stderr.log)\tFAIL' >> /test_output/test_results.tsv \
|| echo -e 'No sanitizer asserts\tOK' >> /test_output/test_results.tsv
rm -f /test_output/tmp
# OOM
zgrep -Fa " <Fatal> Application: Child process was terminated by signal 9" /var/log/clickhouse-server/clickhouse-server.log* > /dev/null \
&& echo -e 'OOM killer (or signal 9) in clickhouse-server.log\tFAIL' >> /test_output/test_results.tsv \
|| echo -e 'No OOM messages in clickhouse-server.log\tOK' >> /test_output/test_results.tsv
# Logical errors
zgrep -Fa "Code: 49, e.displayText() = DB::Exception:" /var/log/clickhouse-server/clickhouse-server.log* > /dev/null \
&& echo -e 'Logical error thrown (see clickhouse-server.log)\tFAIL' >> /test_output/test_results.tsv \
|| echo -e 'No logical errors\tOK' >> /test_output/test_results.tsv
# Crash
zgrep -Fa "########################################" /var/log/clickhouse-server/clickhouse-server.log* > /dev/null \
&& echo -e 'Killed by signal (in clickhouse-server.log)\tFAIL' >> /test_output/test_results.tsv \
|| echo -e 'Not crashed\tOK' >> /test_output/test_results.tsv
# It also checks for crash without stacktrace (printed by watchdog)
zgrep -Fa " <Fatal> " /var/log/clickhouse-server/clickhouse-server.log* > /dev/null \
&& echo -e 'Fatal message in clickhouse-server.log\tFAIL' >> /test_output/test_results.tsv \
|| echo -e 'No fatal messages in clickhouse-server.log\tOK' >> /test_output/test_results.tsv
zgrep -Fa "########################################" /test_output/* > /dev/null \
&& echo -e 'Killed by signal (output files)\tFAIL' >> /test_output/test_results.tsv
zgrep -Fa " received signal " /test_output/gdb.log > /dev/null \
&& echo -e 'Found signal in gdb.log\tFAIL' >> /test_output/test_results.tsv
# Put logs into /test_output/
for log_file in /var/log/clickhouse-server/clickhouse-server.log*
do
pigz < "${log_file}" > /test_output/"$(basename ${log_file})".gz
# FIXME: remove once only github actions will be left
rm "${log_file}"
done
tar -chf /test_output/coordination.tar /var/lib/clickhouse/coordination ||:
mv /var/log/clickhouse-server/stderr.log /test_output/
# Replace the engine with Ordinary to avoid extra symlinks stuff in artifacts.
# (so that clickhouse-local --path can read it w/o extra care).
sed -i -e "s/ATTACH DATABASE _ UUID '[^']*'/ATTACH DATABASE system/" -e "s/Atomic/Ordinary/" /var/lib/clickhouse/metadata/system.sql
for table in query_log trace_log; do
sed -i "s/ATTACH TABLE _ UUID '[^']*'/ATTACH TABLE $table/" /var/lib/clickhouse/metadata/system/${table}.sql
tar -chf /test_output/${table}_dump.tar /var/lib/clickhouse/metadata/system.sql /var/lib/clickhouse/metadata/system/${table}.sql /var/lib/clickhouse/data/system/${table} ||:
done
# Write check result into check_status.tsv
clickhouse-local --structure "test String, res String" -q "SELECT 'failure', test FROM table WHERE res != 'OK' order by (lower(test) like '%hung%') LIMIT 1" < /test_output/test_results.tsv > /test_output/check_status.tsv
[ -s /test_output/check_status.tsv ] || echo -e "success\tNo errors found" > /test_output/check_status.tsv
# Core dumps (see gcore)
# Default filename is 'core.PROCESS_ID'
for core in core.*; do
pigz $core
mv $core.gz /output/
done