7d18ea5e93
When running in verbosity <2, display the file and line number for tasks that fail. This provides useful information without having to run at increased verbosity. * Move _print_task_path to CallbackBase class * Add integration tests * Add color parameter to _print_task_path() * Keep color output consistent for now Currently the path is display with COLOR_DEBUG formatting with verbosity >= 2. Instead of the color of the path changing based on verbosity level, just keep it at the currently behavior of COLOR_DEBUG. Having the color of the same information change based on verbosity level seems incorrect and makes visual parsing of the information more difficult. Co-authored-by: tahar.jegham <jeghamseifeddine@gmail.com>
207 lines
6.2 KiB
Bash
Executable file
207 lines
6.2 KiB
Bash
Executable file
#!/usr/bin/env bash
|
|
|
|
# This test compares "known good" output with various settings against output
|
|
# with the current code. It's brittle by nature, but this is probably the
|
|
# "best" approach possible.
|
|
#
|
|
# Notes:
|
|
# * options passed to this script (such as -v) are ignored, as they would change
|
|
# the output and break the test
|
|
# * the number of asterisks after a "banner" differs depending on the number of
|
|
# columns on the TTY, so we must adjust the columns for the current session
|
|
# for consistency
|
|
|
|
set -eux
|
|
|
|
run_test() {
|
|
local testname=$1
|
|
|
|
# output was recorded w/o cowsay, ensure we reproduce the same
|
|
export ANSIBLE_NOCOWS=1
|
|
|
|
# The shenanigans with redirection and 'tee' are to capture STDOUT and
|
|
# STDERR separately while still displaying both to the console
|
|
{ ansible-playbook -i inventory test.yml \
|
|
> >(set +x; tee "${OUTFILE}.${testname}.stdout"); } \
|
|
2> >(set +x; tee "${OUTFILE}.${testname}.stderr" >&2)
|
|
# Scrub deprication warning that shows up in Python 2.6 on CentOS 6
|
|
sed -i -e '/RandomPool_DeprecationWarning/d' "${OUTFILE}.${testname}.stderr"
|
|
sed -i -e 's/included: .*\/test\/integration/included: ...\/test\/integration/g' "${OUTFILE}.${testname}.stdout"
|
|
sed -i -e 's/@@ -1,1 +1,1 @@/@@ -1 +1 @@/g' "${OUTFILE}.${testname}.stdout"
|
|
sed -i -e 's/: .*\/test_diff\.txt/: ...\/test_diff.txt/g' "${OUTFILE}.${testname}.stdout"
|
|
sed -i -e "s#${ANSIBLE_PLAYBOOK_DIR}#TEST_PATH#g" "${OUTFILE}.${testname}.stdout"
|
|
|
|
diff -u "${ORIGFILE}.${testname}.stdout" "${OUTFILE}.${testname}.stdout" || diff_failure
|
|
diff -u "${ORIGFILE}.${testname}.stderr" "${OUTFILE}.${testname}.stderr" || diff_failure
|
|
}
|
|
|
|
run_test_dryrun() {
|
|
local testname=$1
|
|
# optional, pass --check to run a dry run
|
|
local chk=${2:-}
|
|
|
|
# outout was recorded w/o cowsay, ensure we reproduce the same
|
|
export ANSIBLE_NOCOWS=1
|
|
|
|
# This needed to satisfy shellcheck that can not accept unquoted variable
|
|
cmd="ansible-playbook -i inventory ${chk} test_dryrun.yml"
|
|
|
|
# The shenanigans with redirection and 'tee' are to capture STDOUT and
|
|
# STDERR separately while still displaying both to the console
|
|
{ $cmd \
|
|
> >(set +x; tee "${OUTFILE}.${testname}.stdout"); } \
|
|
2> >(set +x; tee "${OUTFILE}.${testname}.stderr" >&2)
|
|
# Scrub deprication warning that shows up in Python 2.6 on CentOS 6
|
|
sed -i -e '/RandomPool_DeprecationWarning/d' "${OUTFILE}.${testname}.stderr"
|
|
|
|
diff -u "${ORIGFILE}.${testname}.stdout" "${OUTFILE}.${testname}.stdout" || diff_failure
|
|
diff -u "${ORIGFILE}.${testname}.stderr" "${OUTFILE}.${testname}.stderr" || diff_failure
|
|
}
|
|
|
|
diff_failure() {
|
|
if [[ $INIT = 0 ]]; then
|
|
echo "FAILURE...diff mismatch!"
|
|
exit 1
|
|
fi
|
|
}
|
|
|
|
cleanup() {
|
|
if [[ $INIT = 0 ]]; then
|
|
rm -rf "${OUTFILE}.*"
|
|
fi
|
|
|
|
if [[ -f "${BASEFILE}.unreachable.stdout" ]]; then
|
|
rm -rf "${BASEFILE}.unreachable.stdout"
|
|
fi
|
|
|
|
if [[ -f "${BASEFILE}.unreachable.stderr" ]]; then
|
|
rm -rf "${BASEFILE}.unreachable.stderr"
|
|
fi
|
|
|
|
# Restore TTY cols
|
|
if [[ -n ${TTY_COLS:-} ]]; then
|
|
stty cols "${TTY_COLS}"
|
|
fi
|
|
}
|
|
|
|
adjust_tty_cols() {
|
|
if [[ -t 1 ]]; then
|
|
# Preserve existing TTY cols
|
|
TTY_COLS=$( stty -a | grep -Eo '; columns [0-9]+;' | cut -d';' -f2 | cut -d' ' -f3 )
|
|
# Override TTY cols to make comparing ansible-playbook output easier
|
|
# This value matches the default in the code when there is no TTY
|
|
stty cols 79
|
|
fi
|
|
}
|
|
|
|
BASEFILE=callback_default.out
|
|
|
|
ORIGFILE="${BASEFILE}"
|
|
OUTFILE="${BASEFILE}.new"
|
|
|
|
trap 'cleanup' EXIT
|
|
|
|
# The --init flag will (re)generate the "good" output files used by the tests
|
|
INIT=0
|
|
if [[ ${1:-} == "--init" ]]; then
|
|
shift
|
|
OUTFILE=$ORIGFILE
|
|
INIT=1
|
|
fi
|
|
|
|
adjust_tty_cols
|
|
|
|
# Force the 'default' callback plugin, since that's what we're testing
|
|
export ANSIBLE_STDOUT_CALLBACK=default
|
|
# Disable color in output for consistency
|
|
export ANSIBLE_FORCE_COLOR=0
|
|
export ANSIBLE_NOCOLOR=1
|
|
|
|
# Default settings
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=1
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=1
|
|
export ANSIBLE_DISPLAY_FAILED_STDERR=0
|
|
export ANSIBLE_CHECK_MODE_MARKERS=0
|
|
|
|
run_test default
|
|
|
|
# Hide skipped
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=0
|
|
|
|
run_test hide_skipped
|
|
|
|
# Hide skipped/ok
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=0
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=0
|
|
|
|
run_test hide_skipped_ok
|
|
|
|
# Hide ok
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=1
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=0
|
|
|
|
run_test hide_ok
|
|
|
|
# Failed to stderr
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=1
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=1
|
|
export ANSIBLE_DISPLAY_FAILED_STDERR=1
|
|
|
|
run_test failed_to_stderr
|
|
export ANSIBLE_DISPLAY_FAILED_STDERR=0
|
|
|
|
|
|
# Test displaying task path on failure
|
|
export ANSIBLE_SHOW_TASK_PATH_ON_FAILURE=1
|
|
run_test display_path_on_failure
|
|
export ANSIBLE_SHOW_TASK_PATH_ON_FAILURE=0
|
|
|
|
|
|
# Default settings with unreachable tasks
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=1
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=1
|
|
export ANSIBLE_DISPLAY_FAILED_STDERR=1
|
|
export ANSIBLE_TIMEOUT=1
|
|
|
|
# Check if UNREACHBLE is available in stderr
|
|
set +e
|
|
ansible-playbook -i inventory test_2.yml > >(set +x; tee "${BASEFILE}.unreachable.stdout";) 2> >(set +x; tee "${BASEFILE}.unreachable.stderr" >&2) || true
|
|
set -e
|
|
if test "$(grep -c 'UNREACHABLE' "${BASEFILE}.unreachable.stderr")" -ne 1; then
|
|
echo "Test failed"
|
|
exit 1
|
|
fi
|
|
|
|
## DRY RUN tests
|
|
#
|
|
# Default settings with dry run tasks
|
|
export ANSIBLE_DISPLAY_SKIPPED_HOSTS=1
|
|
export ANSIBLE_DISPLAY_OK_HOSTS=1
|
|
export ANSIBLE_DISPLAY_FAILED_STDERR=1
|
|
# Enable Check mode markers
|
|
export ANSIBLE_CHECK_MODE_MARKERS=1
|
|
|
|
# Test the wet run with check markers
|
|
run_test_dryrun check_markers_wet
|
|
|
|
# Test the dry run with check markers
|
|
run_test_dryrun check_markers_dry --check
|
|
|
|
# Disable Check mode markers
|
|
export ANSIBLE_CHECK_MODE_MARKERS=0
|
|
|
|
# Test the wet run without check markers
|
|
run_test_dryrun check_nomarkers_wet
|
|
|
|
# Test the dry run without check markers
|
|
run_test_dryrun check_nomarkers_dry --check
|
|
|
|
# Make sure implicit meta tasks are not printed
|
|
ansible-playbook -i host1,host2 no_implicit_meta_banners.yml > meta_test.out
|
|
cat meta_test.out
|
|
[ "$(grep -c 'TASK \[meta\]' meta_test.out)" -eq 0 ]
|
|
rm -f meta_test.out
|
|
|
|
# Ensure free/host_pinned non-lockstep strategies display correctly
|
|
diff -u callback_default.out.free.stdout <(ANSIBLE_STRATEGY=free ansible-playbook -i inventory test_non_lockstep.yml 2>/dev/null)
|
|
diff -u callback_default.out.host_pinned.stdout <(ANSIBLE_STRATEGY=host_pinned ansible-playbook -i inventory test_non_lockstep.yml 2>/dev/null)
|