proxysql

Commit Graph

Author	SHA1	Message	Date
Rene Cannao	99c86e5759	test: run check_groups.py at start of run-multi-group.bash Abort CI early if any compiled test binaries are missing from groups.json. This prevents silent test omissions — if a developer adds a test but forgets to register it, CI fails immediately with a clear message instead of silently skipping it.	1 month ago
Rene Cannao	39840b5393	Exclude /usr and deps from coverage reports fastcov was including system headers (/usr/include/c++/...) and vendored dependencies (deps/json/json.hpp) in coverage output. Use --exclude instead of --include since project source files have bare filenames in gcov output (no path prefix to match against), while system/deps paths are always absolute.	1 month ago
Rene Cannao	5a8b6e9a89	Fix gcov coverage collection for multi-group CI runs Three problems caused machine stalls and 0% coverage in multi-group runs: 1. All test-runner containers wrote TAP test .gcda files to the shared ${WORKSPACE}, causing concurrent corruption and memory bloat when multiple gcov processes fought over the same files. 2. Each container independently ran fastcov -j$(nproc) in its EXIT trap, spawning hundreds of concurrent gcov processes across all containers. 3. GCOV_PREFIX_STRIP=3 placed .gcda at proxysql/lib/obj/ but .gcno was copied to lib/obj/ — path mismatch meant gcov couldn't correlate them. Additionally, the --include filter expected full workspace paths but gcov records bare filenames, so all results were filtered out. Changes: - Set GCOV_PREFIX=/gcov/tap on test-runner containers so TAP .gcda files go to per-INFRA_ID /gcov/tap/ instead of shared workspace. - Pass MULTI_GROUP=1 from run-multi-group.bash to test-runners. When set, the EXIT trap copies .gcno adjacent to .gcda (handling path offset from GCOV_PREFIX_STRIP) but skips fastcov/genhtml. - Add centralized coverage collection in run-multi-group.bash: processes each group's /gcov directory sequentially with fastcov -j4 after all groups finish, then combines the .info files. - Remove --include filter from fastcov (source paths are bare filenames, not full paths, so the filter rejected everything). - Standalone mode (no MULTI_GROUP) is unchanged.	1 month ago
Rene Cannao	83c3e76506	Fix libparser.so loading for CI test binaries TAP test binaries link libcpp_dotenv.so which transitively needs libparser.so. The RUNPATH baked into libcpp_dotenv.so contains stale paths from the original build, so the dynamic linker can't find it. - Add LD_LIBRARY_PATH for the SKIP_PROXYSQL (unit test) host path - Add TAP_DEPS_PATH to env-isolated.bash so proxysql-tester.py constructs the correct LD_LIBRARY_PATH inside Docker containers	1 month ago
Rene Cannao	f8c64142cb	Fix cluster CI: remove duplicate primary INSERT and add sync_algorithm=3 - Remove redundant INSERT of primary into proxysql_servers on nodes. Commit `b02147b92` added the primary to the shared PROXYSQL_SERVERS_SQL but didn't remove the pre-existing per-node INSERT, causing a UNIQUE constraint violation that broke all multi-node CI groups. - Add cluster_mysql_servers_sync_algorithm=3 to test_cluster_sync config files and update hardcoded datadir paths to match current workspace.	1 month ago
Rene Cannao	c7647c7387	Fix coverage: generate separate reports and combine with fastcov -C Coverage data comes from two sources that may overlap: /gcov — ProxySQL daemon execution (via GCOV_PREFIX) WORKSPACE — TAP test execution (.gcda next to binaries) Both may cover the same code (e.g. libproxysql.a exercised by both ProxySQL runtime and TAP test harness). Previously we copied .gcda files between directories, which either overwrote (cp -f) or kept only the first version (cp -n) instead of combining counters. Now generates two separate .info files (proxysql.info + tap_tests.info) and combines them with fastcov -C, which correctly sums execution counters for overlapping source files.	1 month ago
Rene Cannao	8ca54a9def	Fix coverage: don't overwrite ProxySQL's gcda files with stale copies The coverage collection copied ALL .gcno and .gcda files from the workspace to /gcov using cp -f. But ProxySQL's .gcda files (with real execution data from the Docker container) were already in /gcov via GCOV_PREFIX. The cp -f overwrote them with the host's stale .gcda files (zero or outdated execution counts). This caused lib/MySQLFFTO.cpp and lib/PgSQLFFTO.cpp (and likely other ProxySQL library files) to show 0% coverage despite the FFTO tests running successfully and the .gcda files having real data. Fix: copy .gcno files with cp -f (always needed for fastcov), but copy .gcda files with cp -n (no-clobber) so ProxySQL's existing execution data in /gcov is preserved. TAP test .gcda files that don't exist in /gcov yet are still copied.	1 month ago
Rene Cannao	b02147b92b	Include primary in its own proxysql_servers table When a cluster node syncs proxysql_servers from the primary and runs LOAD PROXYSQL SERVERS TO RUNTIME, it rebuilds the node list from the synced data. If the primary isn't in its own proxysql_servers table, the synced list won't include it. The node then destroys the cluster monitor thread for the primary and never recreates it. Evidence: node 1 log shows "Destroyed Cluster Node Entry for host proxysql:6032" followed by "closing thread for peer proxysql:6032" after syncing proxysql_servers from the primary. After that, the node never detected any further checksum changes from the primary, causing test_cluster1-t and test_unshun_algorithm-t to fail. Fixes #5542.	1 month ago
Rene Cannao	3c4ebf0da7	Set cluster_mysql_servers_sync_algorithm=3 (auto_select) on all nodes The default value (1) syncs both configured AND runtime mysql_servers, including changes made by the monitor (read_only checks moving servers between writer/reader hostgroups). This caused circular sync: a cluster node's monitor modified servers, creating a newer epoch, and the primary synced this back — overwriting its correct config and saving the corrupted version to disk. With auto_select (3), the algorithm chooses mysql_servers_v2 when the monitor is enabled, which only syncs user-configured changes and ignores monitor-triggered runtime modifications. Evidence: node 6042's monitor removed MariaDB servers from hostgroup 1701 ("Setting status OFFLINE HARD"), creating checksum 0x7C0A67D5 with epoch 1774288555. Primary (epoch 1774288549) synced from node 6042 and saved to disk, losing hostgroup 1300 permanently. Fixes #5540.	1 month ago
Rene Cannao	fb9ff4476e	Include TAP test coverage: copy .gcda files from source tree to /gcov TAP test binaries are compiled with gcov but their .gcda files are written to the source tree (next to the binaries), not to /gcov. The coverage collection only scanned /gcov, missing all TAP test coverage data. Now copies both .gcno and .gcda files from the workspace into /gcov before running fastcov, so TAP test coverage is included in reports.	1 month ago
Rene Cannao	0f12f5d6f1	Fix TAP_CLUSTER_NODES: include primary (6032) as first entry test_cluster1-t expects conns[0] to be the primary ProxySQL instance. TAP_CLUSTER_NODES only listed nodes 1-9 (ports 6042-6122), so the test loop operated on nodes 1-4 instead of primary + nodes 1-3. Node 4 (port 6072) is NOT in proxysql_servers (only core nodes 1-3 are registered), so changes on node 4 were invisible to the cluster — sync could never detect them, causing the 30-second timeout. Now TAP_CLUSTER_NODES starts with proxysql:6032 (primary), matching the old cluster_ports layout where 6032 was the first entry. Fixes #5535.	1 month ago
Rene Cannao	a4675096b5	Fix cluster admin_variables sync: set same admin vars on primary and nodes The primary had admin-debug=false and admin-restapi_enabled=false (defaults), while cluster nodes had them set to true. Since these variables are NOT in the CLUSTER_SYNC_INTERFACES_ADMIN exclusion list, they're included in the checksum. The checksum mismatch between primary and nodes caused admin_variables sync to never converge on satellite nodes. Now the primary sets the same admin-restapi_enabled=true and admin-debug=true as all nodes, ensuring matching checksums. Fixes #5535.	1 month ago
Rene Cannao	82c112c461	Fix coverage: restore GCOV_PREFIX for primary after unsetting for cluster nodes	1 month ago
Rene Cannao	c84f4117ac	Fix pgsql admin port: set pgsql_ifaces in admin_variables, not pgsql_variables	1 month ago
Rene Cannao	98b56ec386	Move cluster node pgsql/sqlite ports to 7xxx range to avoid restapi collisions	1 month ago
Rene Cannao	48376427cf	Fix port conflicts: assign unique pgsql/sqlite ports to cluster nodes Cluster nodes only configured mysql admin and data ports, leaving pgsql (6133/6132) and sqlite (6030) at defaults — same as the primary. All nodes tried to bind the same ports, causing: 1. Bind failures for pgsql on cluster nodes 2. Cluster sync overwriting primary's pgsql config with empty tables (nodes had no working pgsql, so their pgsql tables were empty) Now each node gets unique ports for all interfaces: node N: mysql admin=6032+N10, mysql data=6033+N10, pgsql data=6133+N10, pgsql admin=6132+N10, sqlite=6030+N*10	1 month ago
Rene Cannao	35a9bfb42e	Use 'proxysql' hostname in proxysql_servers instead of 127.0.0.1 TAP tests running in the test-runner container query proxysql_servers to find cluster nodes, then connect to them directly. With 127.0.0.1, the test-runner connects to itself instead of the ProxySQL container. Using 'proxysql' (the container hostname) works from both inside the container (resolves via /etc/hosts) and from the test-runner (resolves via Docker DNS on the shared network).	1 month ago
Rene Cannao	8705685f83	Address PR review comments: port collisions, gcov, loop bounds, fastcov - RESTAPI port offset changed from 6070+i to 7070+i to avoid collision with admin/mysql ports (6072 would collide with node4 admin) - Unset GCOV_PREFIX for background cluster nodes to prevent concurrent .gcda write corruption - Cap core-node and scheduler loops to min(NUM_NODES, 3) so smaller clusters don't reference non-existent nodes - TAP_CLUSTER_NODES now respects PROXYSQL_CLUSTER_NODES and SKIP_CLUSTER_START instead of hardcoding 9 - Fix fastcov flag: -i is invalid, use --include for source filtering	1 month ago
Rene Cannao	5a868d3366	Default AUTO_CLEANUP=1 in run-multi-group.bash	1 month ago
Rene Cannao	ce92a2a0e4	Redirect coverage generation output to log file in run-multi-group.bash fastcov, lcov, and genhtml produce verbose output that clutters the multi-group summary. Now redirected to coverage-generation.log inside the combined coverage report directory. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	0590803889	Limit code coverage to proxysql source directories only Replace exclude-based filtering (-e /usr/include, deps/) with include-based filtering (-i) for include/, lib/, src/, test/. This avoids capturing coverage for system headers, vendored dependencies, and other non-proxysql code. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	8c67073c15	Redirect coverage generation output to dedicated log file fastcov and genhtml produce verbose output that clutters the group test logs. Now redirected to coverage-generation.log inside the coverage report directory. Status messages still go to stdout. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	6fcb8fd631	Run ProxySQL cluster nodes inside a single container Instead of spawning 1+9 Docker containers for the ProxySQL cluster, run all instances as background processes inside the primary container. Each node gets its own port pair (admin=6032+i10, mysql=6033+i10) and data directory. This eliminates 9 Docker containers per group, avoiding Docker networking/iptables contention that caused "no route to host" errors under parallel test execution. Changes: - start-proxysql-isolated.bash: starts cluster nodes as background processes inside the container, then initializes the cluster and installs the scheduler — all in one script - check_all_nodes.bash: polls 127.0.0.1 on different ports instead of container hostnames - env-isolated.bash: TAP_CLUSTER_NODES uses proxysql:PORT format - default/pre-proxysql.bash, legacy/pre-proxysql.bash: no longer call cluster_start.bash/cluster_init.bash (handled by start-proxysql) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	879ebdc220	Fix startup order and add readiness check in test runner ensure-infras.bash: Move pre-proxysql hook (cluster setup) to run BEFORE backend infras start. Backends need ProxySQL and cluster to be fully ready because their docker-proxy-post.bash configures ProxySQL. Previous order was: ProxySQL → infras → cluster. New order: ProxySQL → cluster → infras (one by one) → setup hook. run-tests-isolated.bash: Add a readiness loop that waits up to 30s for ProxySQL admin port to be reachable from inside the test-runner container before running anything. Docker DNS resolution on newly created containers can take several seconds under load, causing "no route to host" (errno 113) errors on the first connection attempt. On failure, dumps DNS resolution diagnostics instead of silently proceeding. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	7df23de546	Remove error suppression from ProxySQL config dump queries Stop redirecting stderr to /dev/null and silently continuing on failure. If the config dump queries fail, we need to see the actual error and the test run should fail visibly, not silently proceed with a broken configuration. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	98fb842c50	Add test/tap/tests/unit/ to TAP_WORKDIRS Unit test binaries live in test/tap/tests/unit/ but TAP_WORKDIRS only included test/tap/tests/ and tests_with_deps/. Without this, proxysql-tester.py cannot find unit test binaries. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	4dfaedcb84	Revert "Skip missing unit test binaries instead of counting as failures" This reverts commit `75d324680d`.	1 month ago
Rene Cannao	75d324680d	Skip missing unit test binaries instead of counting as failures Unit tests in groups.json may reference binaries that haven't been compiled (they require a separate build step). Previously, not-found binaries were counted as test failures. Now they are skipped with a message, so the group passes if all compiled tests pass. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	6d90bc2a7e	Source group env.sh inside test runner container The group's env.sh (setting TEST_PY_* flags) was only sourced on the host. The Docker test runner container didn't see these values, so env-isolated.bash defaults took over (all tests disabled). Now source group env.sh before env-isolated.bash inside the container so group-specific settings are preserved. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
Rene Cannao	3389217e98	Default WITHGCOV=0 in env-isolated.bash Code coverage should be opt-in, not default. Non-coverage builds fail with 'PROXYSQL GCOV DUMP' syntax error when WITHGCOV defaults to 1. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	1 month ago
René Cannaò	3aab0c280f	Fix SKIP_PROXYSQL path: run test binaries directly instead of proxysql-tester.py proxysql-tester.py requires a ProxySQL admin connection and Docker-specific env vars (TESTS_LOGS_PATH, TAP_ADMINPORT, TAP_WORKDIR, etc.) that don't exist when SKIP_PROXYSQL=1. Instead, discover test binaries from groups.json and execute them directly on the host with a simple pass/fail summary. Addresses review feedback from CodeRabbit, Copilot, and Gemini on PR #5520.	1 month ago
René Cannaò	4c4736c73d	Add SKIP_PROXYSQL support for infrastructure-free test groups Unit tests run standalone against libproxysql.a with stubs and need no ProxySQL instance or backend databases. This adds a SKIP_PROXYSQL flag that test groups can set in their env.sh to bypass all Docker infrastructure setup and run tests directly on the host. Changes: - ensure-infras.bash: source group env.sh early; exit immediately when SKIP_PROXYSQL=1 - run-tests-isolated.bash: source group env.sh early; skip ProxySQL and backend verification; run proxysql-tester.py directly on host instead of inside a Docker container - New test/tap/groups/unit-tests/env.sh with SKIP_PROXYSQL=1 - groups.json: add 18 unit test binaries to unit-tests-g1 group - Documentation: SKIP_PROXYSQL in env var tables and new section for infrastructure-free groups in test/infra/README.md	1 month ago
René Cannaò	abd1da861f	Merge pull request #5470 from sysown/private/multi-group-runner Multi-group TAP test runner and ClickHouse test fixes	1 month ago
Rene Cannao	60b5aff5c4	Fix tests not running after infrastructure setup in multi-group runner The heredoc in run-multi-group.bash was causing bash to exit prematurely after ensure-infras.bash completed, without executing the test commands. This was due to shell expansion issues with the heredoc structure. Changes: - Replace heredoc with temp file approach in run-multi-group.bash - Add pre-proxysql hook call to ensure-infras.bash for cluster setup - Create legacy/pre-proxysql.bash for cluster node startup - Update mysql84/pre-proxysql.bash with cluster startup calls - Fix coverage collection to work correctly in parallel execution	1 month ago
Rene Cannao	a1042ce13f	Fix code coverage collection: mount gcov directory in test runner The test runner container needs access to the /gcov directory where ProxySQL writes its coverage data (.gcda files). Without this mount, the coverage collection script could only see .gcno files (compile-time) but not the runtime coverage data from ProxySQL. This fix adds the gcov volume mount to the test runner container, allowing it to merge coverage data from each ProxySQL instance.	1 month ago
Rene Cannao	be290c5aeb	Address PR review feedback on test infrastructure - Remove redundant host-side coverage collection in run-tests-isolated.bash (coverage is already collected via trap inside the container) - Initialize TAP_GROUP, group_has_tests, groups before try block in proxysql-tester.py to avoid UnboundLocalError on early exceptions - Differentiate skipped tests from missing tests in TAP_GROUP validation so intentionally skipped tests don't count as missing	1 month ago
Rene Cannao	86e86c8ce3	Fix parallel coverage collection and improve multi-group runner - Fix code coverage race condition: Each ProxySQL instance now writes .gcda files to its own directory using GCOV_PREFIX, preventing parallel groups from overwriting each other's coverage data - Add TAP_USE_NOISE documentation for noise injection testing - Replace random stagger delay with sequential delay (STAGGER_DELAY) to guarantee proper spacing between group startups - Update README with multi-group runner, coverage, and noise injection docs	1 month ago
Rene Cannao	0bc75cd589	Fix test validation and coverage HTML generation - Fix false positive "missing tests" errors when processing secondary workdirs (e.g., deprecate_eof_support). The validation now only checks tests that actually exist in the current workdir rather than all tests in groups.json for the TAP_GROUP. - Add --ignore-errors negative,source --synthesize-missing flags to genhtml calls in run-tests-isolated.bash to match the multi-group runner and prevent HTML generation failures on coverage data with negative branch counts.	1 month ago
Rene Cannao	7ee1332633	Add --ignore-errors source and --synthesize-missing to genhtml Fixes HTML generation errors for missing system headers like /usr/include/c++/13/bits/predefined_ops.h	1 month ago
Rene Cannao	20fc53e45c	Fix code coverage collection in multi-group test runner 1. Add set -e to run-multi-group.bash heredoc for proper error handling 2. Add EXIT trap in run-tests-isolated.bash to collect coverage on timeout/failure 3. Pass COVERAGE_MODE, COVERAGE_REPORT_DIR, and SCRIPT_DIR env vars to container 4. Fix fastcov invocation to use -e deps/ instead of restrictive -i filters 5. Add per-group HTML report generation 6. Run coverage combination in container for host without tools Coverage is now collected for: - include/ (189 header files) - lib/ (109 library files) - src/ (4 source files) The EXIT trap ensures coverage is collected even when tests timeout.	1 month ago
Rene Cannao	30620f37c4	Revert "Move coverage collection inside Docker container" This reverts commit `e571f08a45`.	1 month ago
Rene Cannao	e571f08a45	Move coverage collection inside Docker container Coverage collection now runs INSIDE the test-runner container immediately after tests complete (whether they succeed, fail, or timeout). This ensures: 1. Coverage data is collected even if the parent script times out 2. The .info file is generated inside the container before it exits 3. No dependency on external docker exec after container stops Also removes the external coverage collection code that was unreachable when timeout killed the subshell.	1 month ago
Rene Cannao	5f25df6c61	Improve test startup staggering with group-based hash	1 month ago
Rene Cannao	dcef810b99	Fix inner_exit_code unbound variable in run-multi-group.bash	1 month ago
Rene Cannao	9256a4c2f2	Fix heredoc variable expansion in run-multi-group.bash	1 month ago
Rene Cannao	0d891b8941	Fix variable escaping in run-multi-group.bash subshell	1 month ago
Rene Cannao	64d8149a53	Add COVERAGE to run-multi-group.bash configuration summary output	1 month ago
Rene Cannao	9f6ba655bf	Add optional code coverage support for isolated test runs	1 month ago
Rene Cannao	40e5d785a6	Change multi-group startup delay from 0-15 to 1-20 seconds	1 month ago
Rene Cannao	cf54bff262	Fix docker-proxy-post.bash to use absolute path for .env sourcing The scripts were using relative path [ -f .env ] which failed when called from test/infra/control/, causing all SQL template variables (PREFIX, WHG, RHG, etc.) to be empty. This resulted in query rules with NULL match_digest patterns and incorrect rule IDs.	1 month ago

1 2

72 Commits (dca01d4eaf32c8363ff14fb4f74a46deb73e11c3)