ClickHouse/docker/test/performance-comparison/README.md

# [draft] Performance comparison test

This is an experimental mode that compares performance of old and new server
side by side. Both servers are run, and the query is executed on one then another,
measuring the times. This setup should remove much of the variability present in
the current performance tests, which only run the new version and compare with
the old results recorded some time in the past.

To interpret the observed results, we build randomization distribution for the
observed difference of median times between old and new server, under the null
hypothesis that the performance distribution is the same (for the details of the
method, see [1]). We consider the observed difference in performance significant,
if it is above 5% and above the 95th percentile of the randomization distribution.
We also consider the test to be unstable, if the observed difference is less than
5%, but the 95th percentile is above 5% -- this means that we are likely to observe
performance differences above 5% more often than in 5% runs, so the test is likely
to have false positives.

### How to Read the Report

The check status summarizes the report in a short text message like `1 faster, 10 unstable`:
* `1 faster` -- how many queries became faster,
* `1 slower` -- how many queries are slower,
* `1 too long` -- how many queries are taking too long to run,
* `1 unstable` -- how many queries have unstable results,
* `1 errors` -- how many errors there are in total. Action is required for every error, this number must be zero. The number of errors includes slower tests, tests that are too long, errors while running the tests and building reports, etc. Please look at the main report page to investigate these errors.

The report page itself constists of a several tables. Some of them always signify errors, e.g. "Run errors" -- the very presence of this table indicates that there were errors during the test, that are not normal and must be fixed. Some tables are mostly informational, e.g. "Test times" -- they reflect normal test results. But if a cell in such table is marked in red, this also means an error, e.g., a test is taking too long to run.

#### Tested Commits
Informational, no action required. Log messages for the commits that are tested. Note that for the right commit, we show nominal tested commit `pull/*/head` and real tested commit `pull/*/merge`, which is generated by GitHub by merging latest master to the `pull/*/head` and which we actually build and test in CI.

#### Error Summary
Action required for every item.

This table summarizes all errors that ocurred during the test. Click the links to go to the description of a particular error.

#### Run Errors
Action required for every item -- these are errors that must be fixed.

The errors that ocurred when running some test queries. For more information about the error, download test output archive and see `test-name-err.log`. To reproduce, see 'How to run' below.

#### Slow on Client
Action required for every item -- these are errors that must be fixed.

This table shows queries that take significantly longer to process on the client than on the server. A possible reason might be sending too much data to the client, e.g., a forgotten `format Null`.

#### Unexpected Query Duration
Action required for every item -- these are errors that must be fixed.

A query is supposed to run longer than 0.1 second. If your query runs faster, increase the amount of processed data to bring the run time above this threshold. You can use a bigger table (e.g. `hits_100m` instead of `hits_10m`), increase a `LIMIT`, make a query single-threaded, and so on. Queries that are too fast suffer from poor stability and precision.

Sometimes you want to test a query that is supposed to complete "instantaneously", i.e. in sublinear time. This might be `count(*)`, or parsing a complicated tuple. It might not be practical or even possible to increase the run time of such queries by adding more data. For such queries there is a specal comparison mode which runs them for a fixed amount of time, instead of a fixed number of iterations like we do normally. This mode is inferior to the normal mode, because the influence of noise and overhead is higher, which leads to less precise and stable results.

If it is impossible to increase the run time of a query and it is supposed to complete "immediately", you have to explicitly mark this in the test. To do so, add a `short` attribute to the query tag in the test file: `<query short="1">...`. The value of the `short` attribute is evaluated as a python expression, and substitutions are performed, so you can write something like `<query short="{column1} = {column2}">select count(*) from table where {column1} > {column2}</query>`, to mark only a particular combination of variables as short.

This table shows queries for which the `short` marking is not consistent with the actual query run time -- i.e., a query runs for a normal time but is marked as `short`, or it runs faster than normal but is not marked as `short`.

#### Partial Queries
Action required for the cells marked in red.

Shows the queries we are unable to run on an old server -- probably because they contain a new function. You should see this table when you add a new function and a performance test for it. Check that the run time and variance are acceptable (run time between 0.1 and 1 seconds, variance below 10%). If not, they will be highlighted in red.

#### Changes in Performance
Action required for the cells marked in red, and some cheering is appropriate for the cells marked in green.

These are the queries for which we observe a statistically significant change in performance. Note that there will always be some false positives -- we try to filter by p < 0.001, and have 2000 queries, so two false positives per run are expected. In practice we have more -- e.g. code layout changed because of some unknowable jitter in compiler internals, so the change we observe is real, but it is a 'false positive' in the sense that it is not directly caused by your changes. If, based on your knowledge of ClickHouse internals, you can decide that the observed test changes are not relevant to the changes made in the tested PR, you can ignore them.

You can find flame graphs for queries with performance changes in the test output archive, in files named as 'my_test_0_Cpu_SELECT 1 FROM....FORMAT Null.left.svg'. First goes the test name, then the query number in the test, then the trace type (same as in `system.trace_log`), and then the server version (left is old and right is new).

#### Unstable Queries
Action required for the cells marked in red.

These are the queries for which we did not observe a statistically significant change in performance, but for which the variance in query performance is very high. This means that we are likely to observe big changes in performance even in the absence of real changes, e.g. when comparing the server to itself. Such queries are going to have bad sensitivity as performance tests -- if a query has, say, 50% expected variability, this means we are going to see changes in performance up to 50%, even when there were no real changes in the code. And because of this, we won't be able to detect changes less than 50% with such a query, which is pretty bad. The reasons for the high variability must be investigated and fixed; ideally, the variability should be brought under 5-10%. 

The most frequent reason for instability is that the query is just too short -- e.g. below 0.1 seconds. Bringing query time to 0.2 seconds or above usually helps.
Other reasons may include:
* using a lot of memory which is allocated differently between servers, so the access time may vary. This may apply to your queries if you have a `Memory` engine table that is bigger than 1 GB. For example, this problem has plagued `arithmetic` and `logical_functions` tests for a long time.
* having some threshold behavior in the query, e.g. you insert to a Buffer table and it is flushed only on some query runs, so you get a much higher time for them.

Investigating the instablility is the hardest problem in performance testing, and we still have not been able to understand the reasons behind the instability of some queries. There are some data that can help you in the performance test output archive. Look for files named 'my_unstable_test_0_SELECT 1...FORMAT Null.{left,right}.metrics.rep'. They contain metrics from `system.query_log.ProfileEvents` and functions from stack traces from `system.trace_log`, that vary significantly between query runs. The second column is array of \[min, med, max] values for the metric. Say, if you see `PerfCacheMisses` there, it may mean that the code being tested has not-so-cache-local memory access pattern that is sensitive to memory layout.

#### Skipped Tests
Informational, no action required.

Shows the tests that were skipped, and the reason for it. Normally it is because the data set required for the test was not loaded, or the test is marked as 'long' -- both cases mean that the test is too big to be ran per-commit.

#### Test Performance Changes
Informational, no action required.

This table summarizes the changes in performance of queries in each test -- how many queries have changed, how many are unstable, and what is the magnitude of the changes.

#### Test Times
Action required for the cells marked in red.

This table shows the run times for all the tests. You may have to fix two kinds of errors in this table:
1) Average query run time is too long -- probalby means that the preparatory steps such as creating the table and filling them with data are taking too long. Try to make them faster.
2) Longest query run time is too long -- some particular queries are taking too long, try to make them faster. The ideal query run time is between 0.1 and 1 s.

#### Metric Changes
No action required.

These are changes in median values of metrics from `system.asynchronous_metrics_log`. These metrics are prone to unexplained variation and you can safely ignore this table unless it's interesting to you for some particular reason (e.g. you want to compare memory usage). There are also graphs of these metrics in the performance test output archive, in the `metrics` folder.

#### Errors while Building the Report
Ask a maintainer for help. These errors normally indicate a problem with testing infrastructure.


### How to Run
Run the entire docker container, specifying PR number (0 for master)
and SHA of the commit to test. The reference revision is determined as a nearest
ancestor testing release tag. It is possible to specify the reference revision and
pull requests (0 for master) manually.

```
docker run --network=host --volume=$(pwd)/workspace:/workspace --volume=$(pwd)/output:/output
    [-e REF_PR={} -e REF_SHA={}]
    -e PR_TO_TEST={} -e SHA_TO_TEST={}
    yandex/clickhouse-performance-comparison
```

Then see the `report.html` in the `output` directory.

There are some environment variables that influence what the test does:
 * `-e CHCP_RUNS` -- the number of runs;
 * `-e CHPC_TEST_GREP` -- the names of the tests (xml files) to run, interpreted
 as a grep pattern.
 * `-e CHPC_LOCAL_SCRIPT` -- use the comparison scripts from the docker container and not from the tested commit.

#### Re-genarate report with your tweaks
From the workspace directory (extracted test output archive):
```
stage=report compare.sh
```
More stages are available, e.g. restart servers or run the tests. See the code.

#### Run a single test on the already configured servers
```
docker/test/performance-comparison/perf.py --host=localhost --port=9000 --runs=1 tests/performance/logical_functions_small.xml
```

#### Run all tests on some custom configuration
Start two servers manually on ports `9001` (old) and `9002` (new). Change to a
new directory to be used as workspace for tests, and try something like this:
```
$ PATH=$PATH:~/ch4/build-gcc9-rel/programs \
    CHPC_TEST_PATH=~/ch3/ch/tests/performance \
    CHPC_TEST_GREP=visit_param \
    stage=run_tests \
    ~/ch3/ch/docker/test/performance-comparison/compare.sh
```
* `PATH` must contain `clickhouse-local` and `clickhouse-client`.
* `CHPC_TEST_PATH` -- path to performance test cases, e.g. `tests/performance`.
* `CHPC_TEST_GREP` -- a filter for which tests to run, as a grep pattern.
* `stage` -- from which execution stage to start. To run the tests, use
  `run_tests` stage.

The tests will run, and the `report.html` will be generated in the current
directory.

More complex setup is possible, but inconvenient and requires some scripting.
See `manual-run.sh` for inspiration.

#### Compare two published releases
Use `compare-releases.sh`. It will download and extract static + dbg + test
packages for both releases, and then call the main comparison script
`compare.sh`, starting from `configure` stage.
```
compare-releaseses.sh 19.16.19.85 20.4.2.9
```


#### Statistical considerations
Generating randomization distribution for medians is tricky. Suppose we have N
runs for each version, and then use the combined 2N run results to make a
virtual experiment. In this experiment, we only have N possible values for
median of each version. This becomes very clear if you sort those 2N runs and
imagine where a window of N runs can be -- the N/2 smallest and N/2 largest
values can never be medians. From these N possible values of
medians, you can obtain (N/2)^2 possible values of absolute median difference.
These numbers are +-1, I'm making an off-by-one error somewhere. So, if your
number of runs is small, e.g. 7, you'll only get 16 possible differences, so
even if you make 100k virtual experiments, the randomization distribution will
have only 16 steps, so you'll get weird effects. So you also have to have
enough runs. You can observe it on real data if you add more output to the
query that calculates randomization distribution, e.g., add a number of unique
median values. Looking even more closely, you can see that the exact
values of medians don't matter, and the randomization distribution for
difference of medians devolves into some kind of ranked test. We could probably
skip all these virtual experiments and calculate the resulting distribution
analytically, but I don't know enough math to do it. It would be something
close to Wilcoxon test distribution.

### References
1\. Box, Hunter, Hunter "Statictics for exprerimenters", p. 78: "A Randomized Design Used in the Comparison of Standard and Modified Fertilizer Mixtures for Tomato Plants."