Files
scylladb/test/alternator/run
Nadav Har'El 8f12ef3628 alternator test: faster stream tests by reducing number of vnodes
The single-node test Scylla run by test/alternator/run uses, as the
default, 256 vnodes. When we have 256 vnodes and two shards, our CDC
implementation produces 512 separate "streams" (called "shards" in
DynamoDB lingo). This causes each of the tests in test_streams.py
which need to read data from the stream to need to do 1024 (!) API
requests (512 calls to GetShardIterator and 512 calls to GetRecords)
which takes significant time - about a second per test.

In this patch, we reduce the number of vnodes to 16. We still have
a non-negligible number of stream "shards" (32) so this part of the
CDC code is still exercised. Moreover, to ensure we still routinely
test the paging feature of DescribeStream (whose default page size
is 100), the patch changes the request to use a Limit of 10, so
paging will still be used to retrieve the list of 32 shards.

The time to run the 27 tests in test_streams.py, on my laptop:
Before this patch: 26 seconds
After this patch:   6 seconds.

Fixes #6979
Message-Id: <20200805093418.1490305-1-nyh@scylladb.com>
2020-08-05 19:34:18 +03:00

181 lines
6.4 KiB
Bash
Executable File

#!/bin/bash
# Exit if any one of the commands below fails
set -e
script_path=$(dirname $(readlink -e $0))
source_path=$script_path/../..
# By default, we take the latest build/*/scylla as the executable:
SCYLLA=${SCYLLA-$(ls -t "$source_path/build/"*"/scylla" | head -1)}
SCYLLA=$(readlink -f "$SCYLLA")
# Below, we need to use python3 and the Cassandra drive to set up the
# authentication credentials expected by some of the tests that check
# authentication. If they are not installed there isn't much point of
# even starting Scylla
if ! python3 -c 'from cassandra.cluster import Cluster' >/dev/null 2>&1
then
echo "Error: python3 and python3-cassandra-driver must be installed to configure Alternator authentication." >&2
exit 1
fi
# Pick a loopback IP address for Scylla to run, in an attempt not to collide
# other concurrent runs of Scylla. CCM uses 127.0.0.<nodenum>, so if we use
# 127.1.*.* which cannot collide with it. Moreover, we'll take the last two
# bytes of the address from the current process - so as to allow multiple
# concurrent runs of this code to use a different address.
SCYLLA_IP=127.1.$(($$ >> 8 & 255)).$(($$ & 255))
echo "Running Scylla on $SCYLLA_IP"
tmp_dir="$(readlink -e ${TMPDIR-/tmp})"/alternator-test-$$
mkdir "$tmp_dir"
# We run the cleanup() function on exit for any reason - successful finish
# of the script, an error (since we have "set -e"), or a signal.
# It ensures that Scylla is killed and its temporary storage directory is
# deleted. It also shows Scylla's output log.
code=17
summary=
cleanup() {
kill -9 $SCYLLA_PROCESS 2>/dev/null || :
echo
echo "Scylla log:"
echo
# we want to cat "$tmp_dir/log", but this can take a long time,
# especially if stdout is piped, and be interrupted. We don't want
# the "rm" below to not happen, so we need to open the file later,
# and cat it later:
exec 3<"$tmp_dir/log"
rm -rf --preserve-root "$tmp_dir"
cat <&3
echo $summary
exit $code
}
trap 'cleanup' EXIT
# Set up SSL certificates needed for "--alternator-https-port=8043"
# to work. We only need to do this if the "--https" option was explicitly
# passed - otherwise the test would not use HTTPS anyway.
alternator_port_option="--alternator-port=8000"
alternator_url="http://$SCYLLA_IP:8000"
for i
do
if [ "$i" = --https ]
then
openssl genrsa 2048 > "$tmp_dir/scylla.key"
openssl req -new -x509 -nodes -sha256 -days 365 -subj "/C=IL/ST=None/L=None/O=None/OU=None/CN=example.com" -key "$tmp_dir/scylla.key" -out "$tmp_dir/scylla.crt"
alternator_port_option="--alternator-https-port=8043"
alternator_url="https://$SCYLLA_IP:8043"
fi
done
# To make things easier for users of "killall", "top", and similar, we want
# the Scylla executable which we run during the test to have a different name
# from manul runs of Scylla. Unfortunately, using "exec -a" to change just
# argv[0] isn't good enough - because killall inspects the actual executable
# filename in /proc/<pid>/stat. So we need to name the executable differently.
# Luckily, using a symbolic link is good enough.
SCYLLA_LINK="$tmp_dir"/test_scylla
ln -s "$SCYLLA" "$SCYLLA_LINK"
"$SCYLLA_LINK" --options-file "$source_path/conf/scylla.yaml" \
--alternator-address $SCYLLA_IP \
$alternator_port_option \
--alternator-enforce-authorization=1 \
--alternator-write-isolation=always_use_lwt \
--alternator-streams-time-window-s=0 \
--developer-mode=1 \
--experimental-features=cdc \
--ring-delay-ms 0 --collectd 0 \
--smp 2 -m 1G \
--overprovisioned --unsafe-bypass-fsync 1 \
--api-address $SCYLLA_IP \
--rpc-address $SCYLLA_IP \
--listen-address $SCYLLA_IP \
--prometheus-address $SCYLLA_IP \
--seed-provider-parameters seeds=$SCYLLA_IP \
--workdir "$tmp_dir" \
--server-encryption-options keyfile="$tmp_dir/scylla.key" \
--server-encryption-options certificate="$tmp_dir/scylla.crt" \
--auto-snapshot 0 \
--skip-wait-for-gossip-to-settle 0 \
--logger-log-level compaction=warn \
--logger-log-level migration_manager=warn \
--num-tokens 16 \
>"$tmp_dir/log" 2>&1 &
SCYLLA_PROCESS=$!
# Set up the the proper authentication credentials needed by the Alternator
# test. This requires connecting to Scylla with CQL - we'll wait up for
# one minute for this to work:
setup_authentication() {
python3 -c 'from cassandra.cluster import Cluster; Cluster(["'$SCYLLA_IP'"]).connect().execute("INSERT INTO system_auth.roles (role, salted_hash) VALUES ('\''alternator'\'', '\''secret_pass'\'')")'
}
# Test that Alternator is serving. Alternator starts after CQL, so we use
# check_alternator() to verify that both are up.
check_alternator() {
python3 -c 'import requests; requests.get("'$alternator_url'/")'
}
echo "Scylla is: $SCYLLA."
echo -n "Booting Scylla..."
ok=
SECONDS=0
while ((SECONDS < 200))
do
sleep 1
echo -n .
if ! kill -0 $SCYLLA_PROCESS 2>/dev/null
then
summary="Error: Scylla failed to boot after $SECONDS seconds."
break
fi
case "`check_alternator 2>&1`" in
*"Connection refused"*)
# This may indicate that Scylla is still booting, or that it failed
# to boot and exited. Try again (and check again if Scylla exited).
continue;;
esac
err=`setup_authentication 2>&1` && ok=yes && break
case "$err" in
*NoHostAvailable:*)
# This is what we expect while Scylla is still booting.
;;
*ImportError:*|*"command not found"*)
summary="Error: need python3 and python3-cassandra-driver to configure Alternator authentication."
echo
echo $summary
break;;
*)
summary="Unknown error trying to set authentication credentials: '$err'"
echo
echo $summary
break;;
esac
done
if test -n "$ok" && kill -0 $SCYLLA_PROCESS 2>/dev/null
then
echo "Done ($SECONDS seconds)"
else
echo
if test -z "$summary"
then
summary="Error: Scylla failed to boot after $SECONDS seconds."
echo $summary
fi
exit 1
fi
cd "$script_path"
set +e
pytest --url $alternator_url -o junit_family=xunit2 "$@"
code=$?
case $code in
0) summary="Alternator tests pass";;
*) summary="Alternator tests failure";;
esac
# Note that the cleanup() function runs now, just like on any exit from
# any reason in this script. It will delete the temporary files and
# announce the failure or success of the test.