diff options
author | vkalintiris <vasilis@netdata.cloud> | 2021-10-27 09:26:21 +0300 |
---|---|---|
committer | GitHub <noreply@github.com> | 2021-10-27 09:26:21 +0300 |
commit | 9ed4cea59042aa5e5053c4b4b3529e9d70ab83f9 (patch) | |
tree | a4df8aa963d42f29775224b50059e02a008a1afe /ml/ml.cc | |
parent | a350fb92f921f4765aa73cece3eb23fd8678199b (diff) |
Anomaly Detection MVP (#11548)
* Add support for feature extraction and K-Means clustering.
This patch adds support for performing feature extraction and running the
K-Means clustering algorithm on the extracted features.
We use the open-source dlib library to compute the K-Means clustering
centers, which has been added as a new git submodule.
The build system has been updated to recognize two new options:
1) --enable-ml: build an agent with ml functionality, and
2) --enable-ml-tests: support running tests with the `-W mltest`
option in netdata.
The second flag is meant only for internal use. To build tests successfully,
you need to install the GoogleTest framework on your machine.
* Boilerplate code to track hosts/dims and init ML config options.
A new opaque pointer field is added to the database's host and dimension
data structures. The fields point to C++ wrapper classes that will be used
to store ML-related information in follow-up patches.
The ML functionality needs to iterate all tracked dimensions twice per
second. To avoid locking the entire DB multiple times, we use a
separate dictionary to add/remove dimensions as they are created/deleted
by the database.
A global configuration object is initialized during the startup of the
agent. It will allow our users to specify ML-related configuration
options, eg. hosts/charts to skip from training, etc.
* Add support for training and prediction of dimensions.
Every new host spawns a training thread which is used to train the model
of each dimension.
Training of dimensions is done in a non-batching mode in order to avoid
impacting the generated ML model by the CPU, RAM and disk utilization of
the training code itself.
For performance reasons, prediction is done at the time a new value
is pushed in the database. The alternative option, ie. maintaining a
separate thread for prediction, would be ~3-4x times slower and would
increase locking contention considerably.
For similar reasons, we use a custom function to unpack storage_numbers
into doubles, instead of long doubles.
* Add data structures required by the anomaly detector.
This patch adds two data structures that will be used by the anomaly
detector in follow-up patches.
The first data structure is a circular bit buffer which is being used to
count the number of set bits over time.
The second data structure represents an expandable, rolling window that
tracks set/unset bits. It is explicitly modeled as a finite-state
machine in order to make the anomaly detector's behaviour easier to test
and reason about.
* Add anomaly detection thread.
This patch creates a new anomaly detection thread per host. Each thread
maintains a BitRateWindow which is updated every second based on the
anomaly status of the correspondent host.
Based on the updated status of the anomaly window, we can identify the
existence/absence of an anomaly event, it's start/end time and the
dimensions that participate in it.
* Create/insert/query anomaly events from Sqlite DB.
* Create anomaly event endpoints.
This patch adds two endpoints to expose information about anomaly
events. The first endpoint returns the list of anomalous events within a
specified time range. The second endpoint provides detailed information
about a single anomaly event, ie. the list of anomalous dimensions in
that event along with their anomaly rate.
The `anomaly-bit` option has been added to the `/data` endpoint in order
to allow users to get the anomaly status of individual dimensions per
second.
* Fix build failures on Ubuntu 16.04 & CentOS 7.
These distros do not have toolchains with C++11 enabled by default.
Replacing nullptr with NULL should be fix the build problems on these
platforms when the ML feature is not enabled.
* Fix `make dist` to include ML makefiles and dlib sources.
Currently, we add ml/kmeans/dlib to EXTRA_DIST. We might want to
generate an explicit list of source files in the future, in order to
bring down the generated archive's file size.
* Small changes to make the LGTM & Codacy bots happy.
- Cast unused result of function calls to void.
- Pass a const-ref string to Database's constructor.
- Reduce the scope of a local variable in the anomaly detector.
* Add user configuration option to enable/disable anomaly detection.
* Do not log dimension-specific operations.
Training and prediction operations happen every second for each
dimension. In prep for making this PR easier to run anomaly detection
for many charts & dimensions, I've removed logs that would cause log
flooding.
* Reset dimensions' bit counter when not above anomaly rate threshold.
* Update the default config options with real values.
With this patch the default configuration options will match the ones
we want our users to use by default.
* Update conditions for creating new ML dimensions.
1. Skip dimensions with update_every != 1,
2. Skip dimensions that come from the ML charts.
With this filtering in place, any configuration value for the
relevant simple_pattern expressions will work correctly.
* Teach buildinfo{,json} about the ML feature.
* Set --enable-ml by default in the configuration options.
This patch is only meant for testing the building of the ML functionality
on Github. It will be reverted once tests pass successfully.
* Minor build system fixes.
- Add path to json header
- Enable C++ linker when ML functionality is enabled
- Rename ml/ml-dummy.cc to ml/ml-dummy.c
* Revert "Set --enable-ml by default in the configuration options."
This reverts commit 28206952a59a577675c86194f2590ec63b60506c.
We pass all Github checks when building the ML functionality, except for
those that run on CentOS 7 due to not having a C++11 toolchain.
* Check for missing dlib and nlohmann files.
We simply check the single-source files upon which our build system
depends. If they are missing, an error message notifies the user
about missing git submodules which are required for the ML
functionality.
* Allow users to specify the maximum number of KMeans iterations.
* Use dlib v19.10
v19.22 broke compatibility with CentOS 7's g++. Development of the
anomaly detection used v19.10, which is the version used by most Debian and
Ubuntu distribution versions that are not past EOL.
No observable performance improvements/regressions specific to the K-Means
algorithm occur between the two versions.
* Detect and use the -std=c++11 flag when building anomaly detection.
This patch automatically adds the -std=c++11 when building netdata
with the ML functionality, if it's supported by the user's toolchain.
With this change we are able to build the agent correctly on CentOS 7.
* Restructure configuration options.
- update default values,
- clamp values to min/max defaults,
- validate and identify conflicting values.
* Add update_every configuration option.
Considerring that the MVP does not support per host configuration
options, the update_every option will be used to filter hosts to train.
With this change anomaly detection will be supported on:
- Single nodes with update_every != 1, and
- Children nodes with a common update_every value that might differ from
the value of the parent node.
* Reorganize anomaly detection charts.
This follows Andrew's suggestion to have four charts to show the number
of anomalous/normal dimensions, the anomaly rate, the detector's window
length, and the events that occur in the prediction step.
Context and family values, along with the necessary information in the
dashboard_info.js file, will be updated in a follow-up commit.
* Do not dump anomaly event info in logs.
* Automatically handle low "train every secs" configuration values.
If a user specifies a very low value for the "train every secs", then
it is possible that the time it takes to train a dimension is higher
than the its allotted time.
In that case, we want the training thread to:
- Reduce it's CPU usage per second, and
- Allow the prediction thread to proceed.
We achieve this by limiting the training time of a single dimension to
be equal to half the time allotted to it. This means, that the training
thread will never consume more than 50% of a single core.
* Automatically detect if ML functionality should be enabled.
With these changes, we enable ML if:
- The user has not explicitly specified --disable-ml, and
- Git submodules have been checked out properly, and
- The toolchain supports C++11.
If the user has explicitly specified --enable-ml, the build fails if
git submodules are missing, or the toolchain does not support C++11.
* Disable anomaly detection by default.
* Do not update charts in locked region.
* Cleanup code reading configuration options.
* Enable C++ linker when building ML.
* Disable ML functionality for CMake builds.
* Skip LGTM for dlib and nlohmann libraries.
* Do not build ML if libuuid is missing.
* Fix dlib path in LGTM's yaml config file.
* Add chart to track duration of prediction step.
* Add chart to track duration of training step.
* Limit the number dimensions in an anomaly event.
This will ensure our JSON results won't grow without any limit. The
default ML configuration options, train approximately ~1700 dimensions
in a newly-installed Netdata agent. The hard-limit is set to 2000
dimensions which:
- Is well above the default number of dimensions we train,
- If it is ever reached it means that the user had accidentaly a
very low anomaly rate threshold, and
- Considering that we sort the result by anomaly score, the cutoff
dimensions will be the less anomalous, ie. the least important to
investigate.
* Add information about the ML charts.
* Update family value in ML charts.
This fix will allow us to show the individual charts in the RHS Anomaly
Detection submenu.
* Rename chart type
s/anomalydetection/anomaly_detection/g
* Expose ML feat in /info endpoint.
* Export ML config through /info endpoint.
* Fix CentOS 7 build.
* Reduce the critical region of a host's lock.
Before this change, each host had a single, dedicated lock to protect
its map of dimensions from adding/deleting new dimensions while training
and detecting anomalies. This was problematic because training of a
single dimension can take several seconds in nodes that are under heavy
load.
After this change, the host's lock protects only the insertion/deletion
of new dimensions, and the prediction step. For the training of dimensions
we use a dedicated lock per dimension, which is responsible for protecting
the dimension from deletion while training.
Prediction is fast enough, even on slow machines or under heavy load,
which allows us to use the host's main lock and avoid increasing the
complexity of our implementation in the anomaly detector.
* Improve the way we are tracking anomaly detector's performance.
This change allows us to:
- track the total training time per update_every period,
- track the maximum training time of a single dimension per
update_every period, and
- export the current number of total, anomalous, normal dimensions
to the /info endpoint.
Also, now that we use dedicated locks per dimensions, we can train under
heavy load continuously without having to sleep in order to yield the
training thread and allow the prediction thread to progress.
* Use samples instead of seconds in ML configuration.
This commit changes the way we are handling input ML configuration
options from the user. Instead of treating values as seconds, we
interpret all inputs as number of update_every periods. This allows
us to enable anomaly detection on hosts that have update_every != 1
second, and still produce a model for training/prediction & detection
that behaves in an expected way.
Tested by running anomaly detection on an agent with update_every = [1,
2, 4] seconds.
* Remove unecessary log message in detection thread
* Move ML configuration to global section.
* Update web/gui/dashboard_info.js
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
* Fix typo
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
* Rebase.
* Use negative logic for anomaly bit.
* Add info for prediction_stats and training_stats charts.
* Disable ML on PPC64EL.
The CI test fails with -std=c++11 and requires -std=gnu++11 instead.
However, it's not easy to quickly append the required flag to CXXFLAGS.
For the time being, simply disable ML on PPC64EL and if any users
require this functionality we can fix it in the future.
* Add comment on why we disable ML on PPC64EL.
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
Diffstat (limited to 'ml/ml.cc')
-rw-r--r-- | ml/ml.cc | 153 |
1 files changed, 153 insertions, 0 deletions
diff --git a/ml/ml.cc b/ml/ml.cc new file mode 100644 index 0000000000..857d23d33b --- /dev/null +++ b/ml/ml.cc @@ -0,0 +1,153 @@ +// SPDX-License-Identifier: GPL-3.0-or-later + +#include "Config.h" +#include "Dimension.h" +#include "Host.h" + +using namespace ml; + +/* + * Assumptions: + * 1) hosts outlive their sets, and sets outlive their dimensions, + * 2) dimensions always have a set that has a host. + */ + +void ml_init(void) { + Cfg.readMLConfig(); +} + +void ml_new_host(RRDHOST *RH) { + if (!Cfg.EnableAnomalyDetection) + return; + + if (simple_pattern_matches(Cfg.SP_HostsToSkip, RH->hostname)) + return; + + Host *H = new Host(RH); + RH->ml_host = static_cast<ml_host_t>(H); + + H->startAnomalyDetectionThreads(); +} + +void ml_delete_host(RRDHOST *RH) { + Host *H = static_cast<Host *>(RH->ml_host); + if (!H) + return; + + H->stopAnomalyDetectionThreads(); + + delete H; + RH->ml_host = nullptr; +} + +void ml_new_dimension(RRDDIM *RD) { + RRDSET *RS = RD->rrdset; + + Host *H = static_cast<Host *>(RD->rrdset->rrdhost->ml_host); + if (!H) + return; + + if (static_cast<unsigned>(RD->update_every) != H->updateEvery()) + return; + + if (simple_pattern_matches(Cfg.SP_ChartsToSkip, RS->name)) + return; + + Dimension *D = new Dimension(RD); + RD->state->ml_dimension = static_cast<ml_dimension_t>(D); + H->addDimension(D); +} + +void ml_delete_dimension(RRDDIM *RD) { + Dimension *D = static_cast<Dimension *>(RD->state->ml_dimension); + if (!D) + return; + + Host *H = static_cast<Host *>(RD->rrdset->rrdhost->ml_host); + H->removeDimension(D); + + RD->state->ml_dimension = nullptr; +} + +char *ml_get_host_info(RRDHOST *RH) { + nlohmann::json ConfigJson; + + if (RH && RH->ml_host) { + Host *H = static_cast<Host *>(RH->ml_host); + H->getConfigAsJson(ConfigJson); + H->getDetectionInfoAsJson(ConfigJson); + } else { + ConfigJson["enabled"] = false; + } + + return strdup(ConfigJson.dump(2, '\t').c_str()); +} + +bool ml_is_anomalous(RRDDIM *RD, double Value, bool Exists) { + Dimension *D = static_cast<Dimension *>(RD->state->ml_dimension); + if (!D) + return false; + + D->addValue(Value, Exists); + bool Result = D->predict().second; + return Result; +} + +char *ml_get_anomaly_events(RRDHOST *RH, const char *AnomalyDetectorName, + int AnomalyDetectorVersion, time_t After, time_t Before) { + if (!RH || !RH->ml_host) { + error("No host"); + return nullptr; + } + + Host *H = static_cast<Host *>(RH->ml_host); + std::vector<std::pair<time_t, time_t>> TimeRanges; + + bool Res = H->getAnomaliesInRange(TimeRanges, AnomalyDetectorName, + AnomalyDetectorVersion, + H->getUUID(), + After, Before); + if (!Res) { + error("DB result is empty"); + return nullptr; + } + + nlohmann::json Json = TimeRanges; + return strdup(Json.dump(4).c_str()); +} + +char *ml_get_anomaly_event_info(RRDHOST *RH, const char *AnomalyDetectorName, + int AnomalyDetectorVersion, time_t After, time_t Before) { + if (!RH || !RH->ml_host) { + error("No host"); + return nullptr; + } + + Host *H = static_cast<Host *>(RH->ml_host); + + nlohmann::json Json; + bool Res = H->getAnomalyInfo(Json, AnomalyDetectorName, + AnomalyDetectorVersion, + H->getUUID(), + After, Before); + if (!Res) { + error("DB result is empty"); + return nullptr; + } + + return strdup(Json.dump(4, '\t').c_str()); +} + +#if defined(ENABLE_ML_TESTS) + +#include "gtest/gtest.h" + +int test_ml(int argc, char *argv[]) { + (void) argc; + (void) argv; + + ::testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} + +#endif // ENABLE_ML_TESTS |