Age | Commit message (Collapse) | Author |
|
* Add suppression options.
* Silence noisy dimensions
|
|
|
|
Fix coverity issues 384022, 384021, 383825
|
|
* Pass DB connection in db_execute()
* Add support for loading/saving models.
* Fix ML stats when no training takes place.
* Make model flushing batch size configurable.
* Delete unused function
* Update ML config.
* Restore threshold for logs/period.
* Rm whitespace.
* Add missing dummy function.
* Update function call arguments
* Guard transactions with a lock when flushing ML models.
* Mark dimensions with loaded models as trained.
|
|
* Add acquire/release support for RRDSET
* Release/acquire RRDSET when training.
* Fix function name in log message.
* Use proper function name to get the hostname.
* Add acquire/release for hosts and skip training orphan/obsolete hosts/charts.
* Fix variable name
|
|
This is a temporary change until we figure out
why sqlite degrades the performance of training
threads significantly. To keep the number of
changes to a minimum, a connection is still
opened/closed when initializing/finalizing the
ML component.
|
|
* Apply ML changes again.
The ML changes in 003df5f2 wheere reverted with
556bdad9 because we were partially initializing
ML even when it was explicitly disabled in
netdata.conf, causing the agent to crash on
startup.
* Do not start/stop ML threads when ML is disabled.
* Restore default config settings.
|
|
|
|
* Revert "Revert "Use static thread-pool for training. (#14702)" (#14782)"
This reverts commit 5321ca8d1ef8d974a6a2b2128ca8804de6acb693.
* Model I/O.
* Minor changes
Meant to make debugging a crash issues easier
on cloud VMs:
- Less verbose logging
- Higher logging history
- Modify installer to use debug info by default
* Fix ML initialization order.
* read lock hosts when running detection.
* Revert debugging changes.
* Update ml/Config.cc
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
---------
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
|
|
* configure extent cache size
* workers can now execute up to 10 jobs in a run, boosting query prep and extent reads
* fix dispatched and executing counters
* boost to the max
* increase libuv worker threads
* query prep always get more prio than extent reads; stop processing in batch when dbengine is queue is critical
* fix accounting of query prep
* inlining of time-grouping functions, to speed up queries with billions of points
* make switching based on a local const variable
* print one pending contexts loading message per iteration
* inlined store engine query API
* inlined storage engine data collection api
* inlined all storage engine query ops
* eliminate and inline data collection ops
* simplified query group-by
* more error handling
* optimized partial trimming of group-by queries
* preparative work to support multiple passes of group-by
* more preparative work to support multiple passes of group-by (accepts multiple group-by params)
* unified query timings
* unified query timings - weights endpoint
* query target is no longer a static thread variable - there is a list of cached query targets, each of which of freed every 1000 queries
* fix query memory accounting
* added summary.dimension[].pri and sorted summary.dimensions based on priority and then name
* limit max ACLK WEB response size to 30MB
* the response type should be text/plain
* more preparative work for multiple group-by passes
* create functions for generating group by keys, ids and names
* multiple group-by passes are now supported
* parse group-by options array also with an index
* implemented percentage-of-instance group by function
* family is now merged in multi-node contexts
* prevent uninitialized use
|
|
This reverts commit 5046e034212c008557dd014196b6f6204eda24b2.
Will re-apply once we investigate an issue that occurs during
the shutdown of the agent.
|
|
* Use static thread-pool for training.
* Add missing function definition
* disable training stats chart
* Add config option to explicitly enable ML stats charts.
---------
Co-authored-by: Costa Tsaousis <costa@netdata.cloud>
|
|
* Refactor ML code.
This commit introduces only non-functional changes.
Originally, the C++ code exposed C functions to be called
from the rest of the agent. When we migrated from C++ to C,
we did not eliminate these wrapper functions to make the
PR easier to understand and keep the total LOC low.
This commit removes the wrapper functions and "reclaims" the
`ml_` prefix that we used for the public API of the old
implementation.
Also, the nlohmann Json library has been removed and its
functionality was replaced with the equivalent Json functionality
that we added in libnetdata's BUFFERs.
* Remove missing headers from build systems.
* Fix CMake build.
* rrddim_free is outside of rrd "internals" now.
|
|
* Port ML from C++ to C.
Pretty much everything is a non-functional change, ie.
the functionality is identical to the one provided
by the existing implementation that is written in C++.
Performance-wise, this implementation:
- Eliminates/reduces the number of allocations and
deallocations we have to do for training/detection,
- Uses just a single thread to perform detection for
*all* the hosts (ie. reduces the number of required
threads by 50% on parents), and
- Allows training, prediction and detection of
dimensions that have an update_every that is
different from that of the localhost.
The only C++ functionality that we still use
is vectors, because they make our life easier
and they are pretty much a requirement imposed
by dlib.
* Remove profile.plugin
It was useful only for testing during development.
* Limit logs to 200 lines per period
* Properly generate ml_info in /api/v1/info endpoint.
* Remove resource usage charts since we use worker charts.
* Use a temporary to make linters happy.
* Rebase.
* Fix builds that have ML functionality disabled.
|
|
optimization (#14493)
* first work on standardizing json formatting
* renamed old grouping to time_grouping and added group_by
* add dummy functions to enable compilation
* buffer json api work
* jsonwrap opening with buffer_json_X() functions
* cleanup
* storage for quotes
* optimize buffer printing for both numbers and strings
* removed ; from define
* contexts json generation using the new json functions
* fix buffer overflow at unit test
* weights endpoint using new json api
* fixes to weights endpoint
* check buffer overflow on all buffer functions
* do synchronous queries for weights
* buffer_flush() now resets json state too
* content type typedef
* print double values that are above the max 64-bit value
* str2ndd() can now parse values above UINT64_MAX
* faster number parsing by avoiding double calculations as much as possible
* faster number parsing
* faster hex parsing
* accurate printing and parsing of double values, even for very large numbers that cannot fit in 64bit integers
* full printing and parsing without using library functions - and related unit tests
* added IEEE754 streaming capability to enable streaming of double values in hex
* streaming and replication to transfer all values in hex
* use our own str2ndd for set2
* remove subnormal check from ieee
* base64 encoding for numbers, instead of hex
* when increasing double precision, also make sure the fractional number printed is aligned to the wanted precision
* str2ndd_encoded() parses all encoding formats, including integers
* prevent uninitialized use
* /api/v1/info using the new json API
* Fix error when compiling with --disable-ml
* Remove redundant 'buffer_unittest' declaration
* Fix formatting
* Fix formatting
* Fix formatting
* fix buffer unit test
* apps.plugin using the new JSON API
* make sure the metrics registry does not accept negative timestamps
* do not allow pages with negative timestamps to be loaded from db files; do not accept pages with negative timestamps in the cache
* Fix more formatting
---------
Co-authored-by: Stelios Fragkakis <52996999+stelfrag@users.noreply.github.com>
|
|
* first commit - untested
* fix wrong begin command
* added set v2 too
* debug to log stream buffer
* debug to log stream buffer
* faster streaming printing
* mark charts and dimensions as collected
* use stream points even if sender is not enabled
* comment out stream debug log
* parse null as nan
* custom begin v2
* custom set v2; replication now copies the anomalous flag too
* custom end v2
* enabled stream log test
* renamed to BEGIN2, SET2, END2
* dont mix up replay and v2 members in user object
* fix typo
* cleanup
* support to v2 to v1 proxying
* mark updated dimensions as such
* do not log unknown flags
* comment out stream debug log
* send also the chart id on BEGIN2, v2 to v2
* update the data collections counter
* v2 values are transferred in hex
* faster hex parsing
* a little more generic hex and dec printing and parsing
* fix hex parsing
* minor optimization in dbengine api
* turn debugging into info message
* generalized the timings tracking, so that it can be used in more places
* commented out debug info
* renamed conflicting variable with macro
* remove wrong edits
* integrated ML and added cleanup in case parsing is interrupted
* disable data collection locking during v2
* cleanup stale ML locks; send updated chart variables during v2; add info to find stale locks
* inject an END2 between repeated BEGIN2 from rrdset_done()
* test: remove lockless single-threaded logic from dictionary and aral and apply the right acquire/release memory order to reference counters
* more fine grained dictionary atomics
* remove unecessary return values
* pointer validation under NETDATA_DICTIONARY_VALIDATE_POINTERS
* Revert "pointer validation under NETDATA_DICTIONARY_VALIDATE_POINTERS"
This reverts commit 846cdf2713e2a7ee2ff797f38db11714228800e9.
* Revert "remove unecessary return values"
This reverts commit 8c87d30f4d86f0f5d6b4562cf74fe7447138bbff.
* Revert "more fine grained dictionary atomics"
This reverts commit 984aec4234a340d197d45239ff9a10fd479fcf3c.
* Revert "test: remove lockless single-threaded logic from dictionary and aral and apply the right acquire/release memory order to reference counters"
This reverts commit c460b3d0ad497d2641bd0ea1d63cec7c052e74e4.
* Apply again "pointer validation under NETDATA_DICTIONARY_VALIDATE_POINTERS" while keeping the improved atomic operations.
This reverts commit f158d009
* fix last commit
* fix last commit again
* optimizations in dbengine
* do not send anomaly bit on non-supporting agents (send it when the INTERPOLATED capability is available)
* break long empty-points-loops in rrdset_done()
* decide page alignment on new page allocation, not on every point collected
* create max size pages but no smaller than 1/3
* Fix compilation when --disable-ml is specified
* Return false
* fixes for NETDATA_LOG_REPLICATION_REQUESTS
* added compile option NETDATA_WITHOUT_WORKERS_LATENCY
* put timings in BEGIN2, SET2, END2
* isolate begin2 ml
* revert repositioning data collection lock
* fixed multi-threading of statistics
* do not lookup dimensions all the time if they come in the same order
* update used on iteration, not on every points; also do better error handling
---------
Co-authored-by: Stelios Fragkakis <52996999+stelfrag@users.noreply.github.com>
|
|
* cancel ml threads on shutdown and join them on host free
* mark them not running before joining them
|
|
* count open cache pages refering to datafile
* eliminate waste flush attempts
* remove eliminated variable
* journal v2 scanning split functions
* avoid locking open cache for a long time while migrating to journal v2
* dont acquire datafile for the loop; disable thread cancelability while a query is running
* work on datafile acquiring
* work on datafile deletion
* work on datafile deletion again
* logs of dbengine should start with DBENGINE
* thread specific key for queries to check if a query finishes without a finalize
* page_uuid is not used anymore
* Cleanup judy traversal when building new v2
Remove not needed calls to metric registry
* metric is 8 bytes smaller; timestamps are protected with a spinlock; timestamps in metric are now always coherent
* disable checks for invalid time-ranges
* Remove type from page details
* report scanning time
* remove infinite loop from datafile acquire for deletion
* remove infinite loop from datafile acquire for deletion again
* trace query handles
* properly allocate array of dimensions in replication
* metrics cleanup
* metrics registry uses arrayalloc
* arrayalloc free should be protected by lock
* use array alloc in page cache
* journal v2 scanning fix
* datafile reference leaking hunding
* do not load metrics of future timestamps
* initialize reasons
* fix datafile reference leak
* do not load pages that are entirely overlapped by others
* expand metric retention atomically
* split replication logic in initialization and execution
* replication prepare ahead queries
* replication prepare ahead queries fixed
* fix replication workers accounting
* add router active queries chart
* restore accounting of pages metadata sources; cleanup replication
* dont count skipped pages as unroutable
* notes on services shutdown
* do not migrate to journal v2 too early, while it has pending dirty pages in the main cache for the specific journal file
* do not add pages we dont need to pdc
* time in range re-work to provide info about past and future matches
* finner control on the pages selected for processing; accounting of page related issues
* fix invalid reference to handle->page
* eliminate data collection handle of pg_lookup_next
* accounting for queries with gaps
* query preprocessing the same way the processing is done; cache now supports all operations on Judy
* dynamic libuv workers based on number of processors; minimum libuv workers 8; replication query init ahead uses libuv workers - reserved ones (3)
* get into pdc all matching pages from main cache and open cache; do not do v2 scan if main cache and open cache can satisfy the query
* finner gaps calculation; accounting of overlapping pages in queries
* fix gaps accounting
* move datafile deletion to worker thread
* tune libuv workers and thread stack size
* stop netdata threads gradually
* run indexing together with cache flush/evict
* more work on clean shutdown
* limit the number of pages to evict per run
* do not lock the clean queue for accesses if it is not possible at that time - the page will be moved to the back of the list during eviction
* economies on flags for smaller page footprint; cleanup and renames
* eviction moves referenced pages to the end of the queue
* use murmur hash for indexing partition
* murmur should be static
* use more indexing partitions
* revert number of partitions to number of cpus
* cancel threads first, then stop services
* revert default thread stack size
* dont execute replication requests of disconnected senders
* wait more time for services that are exiting gradually
* fixed last commit
* finer control on page selection algorithm
* default stacksize of 1MB
* fix formatting
* fix worker utilization going crazy when the number is rotating
* avoid buffer full due to replication preprocessing of requests
* support query priorities
* add count of spins in spinlock when compiled with netdata internal checks
* remove prioritization from dbengine queries; cache now uses mutexes for the queues
* hot pages are now in sections judy arrays, like dirty
* align replication queries to optimal page size
* during flushing add to clean and evict in batches
* Revert "during flushing add to clean and evict in batches"
This reverts commit 8fb2b69d068499eacea6de8291c336e5e9f197c7.
* dont lock clean while evicting pages during flushing
* Revert "dont lock clean while evicting pages during flushing"
This reverts commit d6c82b5f40aeba86fc7aead062fab1b819ba58b3.
* Revert "Revert "during flushing add to clean and evict in batches""
This reverts commit ca7a187537fb8f743992700427e13042561211ec.
* dont cross locks during flushing, for the fastest flushes possible
* low-priority queries load pages synchronously
* Revert "low-priority queries load pages synchronously"
This reverts commit 1ef2662ddcd20fe5842b856c716df134c42d1dc7.
* cache uses spinlock again
* during flushing, dont lock the clean queue at all; each item is added atomically
* do smaller eviction runs
* evict one page at a time to minimize lock contention on the clean queue
* fix eviction statistics
* fix last commit
* plain should be main cache
* event loop cleanup; evictions and flushes can now happen concurrently
* run flush and evictions from tier0 only
* remove not needed variables
* flushing open cache is not needed; flushing protection is irrelevant since flushing is global for all tiers; added protection to datafiles so that only one flusher can run per datafile at any given time
* added worker jobs in timer to find the slow part of it
* support fast eviction of pages when all_of_them is set
* revert default thread stack size
* bypass event loop for dispatching read extent commands to workers - send them directly
* Revert "bypass event loop for dispatching read extent commands to workers - send them directly"
This reverts commit 2c08bc5bab12881ae33bc73ce5dea03dfc4e1fce.
* cache work requests
* minimize memory operations during flushing; caching of extent_io_descriptors and page_descriptors
* publish flushed pages to open cache in the thread pool
* prevent eventloop requests from getting stacked in the event loop
* single threaded dbengine controller; support priorities for all queries; major cleanup and restructuring of rrdengine.c
* more rrdengine.c cleanup
* enable db rotation
* do not log when there is a filter
* do not run multiple migration to journal v2
* load all extents async
* fix wrong paste
* report opcodes waiting, works dispatched, works executing
* cleanup event loop memory every 10 minutes
* dont dispatch more work requests than the number of threads available
* use the dispatched counter instead of the executing counter to check if the worker thread pool is full
* remove UV_RUN_NOWAIT
* replication to fill the queues
* caching of extent buffers; code cleanup
* caching of pdc and pd; rework on journal v2 indexing, datafile creation, database rotation
* single transaction wal
* synchronous flushing
* first cancel the threads, then signal them to exit
* caching of rrdeng query handles; added priority to query target; health is now low prio
* add priority to the missing points; do not allow critical priority in queries
* offload query preparation and routing to libuv thread pool
* updated timing charts for the offloaded query preparation
* caching of WALs
* accounting for struct caches (buffers); do not load extents with invalid sizes
* protection against memory booming during replication due to the optimal alignment of pages; sender thread buffer is now also reset when the circular buffer is reset
* also check if the expanded before is not the chart later updated time
* also check if the expanded before is not after the wall clock time of when the query started
* Remove unused variable
* replication to queue less queries; cleanup of internal fatals
* Mark dimension to be updated async
* caching of extent_page_details_list (epdl) and datafile_extent_offset_list (deol)
* disable pgc stress test, under an ifdef
* disable mrg stress test under an ifdef
* Mark chart and host labels, host info for async check and store in the database
* dictionary items use arrayalloc
* cache section pages structure is allocated with arrayalloc
* Add function to wakeup the aclk query threads and check for exit
Register function to be called during shutdown after signaling the service to exit
* parallel preparation of all dimensions of queries
* be more sensitive to enable streaming after replication
* atomically finish chart replication
* fix last commit
* fix last commit again
* fix last commit again again
* fix last commit again again again
* unify the normalization of retention calculation for collected charts; do not enable streaming if more than 60 points are to be transferred; eliminate an allocation during replication
* do not cancel start streaming; use high priority queries when we have locked chart data collection
* prevent starvation on opcodes execution, by allowing 2% of the requests to be re-ordered
* opcode now uses 2 spinlocks one for the caching of allocations and one for the waiting queue
* Remove check locks and NETDATA_VERIFY_LOCKS as it is not needed anymore
* Fix bad memory allocation / cleanup
* Cleanup ACLK sync initialization (part 1)
* Don't update metric registry during shutdown (part 1)
* Prevent crash when dashboard is refreshed and host goes away
* Mark ctx that is shutting down.
Test not adding flushed pages to open cache as hot if we are shutting down
* make ML work
* Fix compile without NETDATA_INTERNAL_CHECKS
* shutdown each ctx independently
* fix completion of quiesce
* do not update shared ML charts
* Create ML charts on child hosts.
When a parent runs a ML for a child, the relevant-ML charts
should be created on the child host. These charts should use
the parent's hostname to differentiate multiple parents that might
run ML for a child.
The only exception to this rule is the training/prediction resource
usage charts. These are created on the localhost of the parent host,
because they provide information specific to said host.
* check new ml code
* first save the database, then free all memory
* dbengine prep exit before freeing all memory; fixed deadlock in cache hot to dirty; added missing check to query engine about metrics without any data in the db
* Cleanup metadata thread (part 2)
* increase refcount before dispatching prep command
* Do not try to stop anomaly detection threads twice.
A separate function call has been added to stop anomaly detection threads.
This commit removes the left over function calls that were made
internally when a host was being created/destroyed.
* Remove allocations when smoothing samples buffer
The number of dims per sample is always 1, ie. we are training and
predicting only individual dimensions.
* set the orphan flag when loading archived hosts
* track worker dispatch callbacks and threadpool worker init
* make ML threads joinable; mark ctx having flushing in progress as early as possible
* fix allocation counter
* Cleanup metadata thread (part 3)
* Cleanup metadata thread (part 4)
* Skip metadata host scan when running unittest
* unittest support during init
* dont use all the libuv threads for queries
* break an infinite loop when sleep_usec() is interrupted
* ml prediction is a collector for several charts
* sleep_usec() now makes sure it will never loop if it passes the time expected; sleep_usec() now uses nanosleep() because clock_nanosleep() misses signals on netdata exit
* worker_unregister() in netdata threads cleanup
* moved pdc/epdl/deol/extent_buffer related code to pdc.c and pdc.h
* fixed ML issues
* removed engine2 directory
* added dbengine2 files in CMakeLists.txt
* move query plan data to query target, so that they can be exposed by in jsonwrap
* uniform definition of query plan according to the other query target members
* event_loop should be in daemon, not libnetdata
* metric_retention_by_uuid() is now part of the storage engine abstraction
* unify time_t variables to have the suffix _s (meaning: seconds)
* old dbengine statistics become "dbengine io"
* do not enable ML resource usage charts by default
* unify ml chart families, plugins and modules
* cleanup query plans from query target
* cleanup all extent buffers
* added debug info for rrddim slot to time
* rrddim now does proper gap management
* full rewrite of the mem modes
* use library functions for madvise
* use CHECKSUM_SZ for the checksum size
* fix coverity warning about the impossible case of returning a page that is entirely in the past of the query
* fix dbengine shutdown
* keep the old datafile lock until a new datafile has been created, to avoid creating multiple datafiles concurrently
* fine tune cache evictions
* dont initialize health if the health service is not running - prevent crash on shutdown while children get connected
* rename AS threads to ACLK[hostname]
* prevent re-use of uninitialized memory in queries
* use JulyL instead of JudyL for PDC operations - to test it first
* add also JulyL files
* fix July memory accounting
* disable July for PDC (use Judy)
* use the function to remove datafiles from linked list
* fix july and event_loop
* add july to libnetdata subdirs
* rename time_t variables that end in _t to end in _s
* replicate when there is a gap at the beginning of the replication period
* reset postponing of sender connections when a receiver is connected
* Adjust update every properly
* fix replication infinite loop due to last change
* packed enums in rrd.h and cleanup of obsolete rrd structure members
* prevent deadlock in replication: replication_recalculate_buffer_used_ratio_unsafe() deadlocking with replication_sender_delete_pending_requests()
* void unused variable
* void unused variables
* fix indentation
* entries_by_time calculation in VD was wrong; restored internal checks for checking future timestamps
* macros to caclulate page entries by time and size
* prevent statsd cleanup crash on exit
* cleanup health thread related variables
Co-authored-by: Stelios Fragkakis <52996999+stelfrag@users.noreply.github.com>
Co-authored-by: vkalintiris <vasilis@netdata.cloud>
|
|
* Add profile.plugin
Creates the specified number of charts/dimensions, and supports
backfilling with pseudo-historical data.
* Bump
* Remove wrongly merged line.
* Use the number of models specified from the config section.
* Add option to consult all ML models.
* Remove profiling option consuming all models.
* Add underscore after chart name prefix.
* prediction -> dimensions chart
* reorder funcs
* Split charts across types with correct priority
* Ignore training request when chart is under replication.
* Track global number of models consulted.
* Cleanup config.
* initial readme updates
* fix readme
* readme
* Fix function definition when ML is disabled.
* Add dummy ml_chart_update_{begin,end}
* Remove profile_plugin
* Define chart priorities under collectors/all.h
* s/curr_t/current_time/
* Use libnetdata's lock/thread wrappers.
* Fix autotools & cmake builds.
* Delete ML dimensions & charts.
* Let users of buffer preprocessing to handle memory.
* Add separate API calls to start/stop ML threads.
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
|
|
(#14172)
|
|
* Add profile.plugin
Creates the specified number of charts/dimensions, and supports
backfilling with pseudo-historical data.
* Bump
* Remove wrongly merged line.
* Use the number of models specified from the config section.
* Add option to consult all ML models.
* Remove profiling option consuming all models.
* Add underscore after chart name prefix.
* prediction -> dimensions chart
* reorder funcs
* Split charts across types with correct priority
* Ignore training request when chart is under replication.
* Track global number of models consulted.
* Cleanup config.
* initial readme updates
* fix readme
* readme
* Fix function definition when ML is disabled.
* Add dummy ml_chart_update_{begin,end}
* Remove profile_plugin
* Define chart priorities under collectors/all.h
* s/curr_t/current_time/
Co-authored-by: Andrew Maguire <andrewm4894@gmail.com>
|
|
|
|
* initial implementation of QUERY_TARGET
* rrd2rrdr() interface
* rrddim_find_best_tier_for_timeframe() ported
* added dimension filtering
* added db object in query target
* rrd2rrdr() ported
* working on formatters
* working on jsonwrapper
* finally, it compiles...
* 1st run without crashes
* query planer working
* cleanup old code
* review changes
* fix also changing data collection frequency
* fix signess
* fix rrdlabels and dimension ordering
* fixes
* remove unused variable
* ml should accept NULL response from rrd2rrdr()
* number formatting fixes
* more number formatting fixes
* more number formatting fixes
* support mc parallel queries
* formatting and cleanup
* added rrd2rrdr_legacy() as a simplified interface to run a query
* make sure rrdset_find_natural_update_every_for_timeframe() returns a value
* make signed comparisons
* weights endpoint using rrdcontexts
* fix for legacy db modes and cleanup
* fix for chart_ids and remove AR chart from weights endpoint
* Ignore command if not initialized yet
* remove unused members
* properly initialize window
* code cleanup - rrddim linked list is gone; rrdset rwlock is gone too
* reviewed RRDR.internal members
* eliminate unnecessary members of QUERY_TARGET
* more complete query ids; more detailed information on aborted queries
* properly terminate option strings
* query id contains group_options which is controlled by users, so escaping is necessary
* tense in query id
* tense in query id - again
* added the remaining query options to the query id
* Expose hidden option to the dimension
* use the hidden flag when loading context dimensions
* Specify table alias for option
* dont update chart last access time, unless at least a dimension of the chart will be queried
Co-authored-by: Stelios Fragkakis <52996999+stelfrag@users.noreply.github.com>
|
|
(#13810)
* overload libc memory allocators with custom ones to trace all allocations
* grab libc pointers for external c plugins
* use -ldl when necessary; fallback to work without dlsym when it is not available
* initialize global variable
* add optional dl libs
* dynamically link every library function when needed for the first time
* prevent crashes on musl libc
* another attempt
* dont dereference function
* attempt no 3
* attempt no 4
* cleanup - all attempts failed
* dont enable tracing of allocations
* missing parenthesis
|
|
* Move all dims under one class.
* Dimension owns anomaly rate RD.
* Remove Dimension::isAnomalous()
* Remove Dimension::trainEvery()
* Rm ml/kmeans
* Remove anomaly detector
The same logic can be implemented by using the host anomaly rate dim.
* Profile plugin.
* Revert "Profile plugin."
This reverts commit e3db37cb49c514502c5216cfe7bca2a003fb90f1.
* Add separate source files for anomaly detection charts.
* Handle training/prediction sync at the dimension level.
* Keep multiple KMeans models in mem.
* Move feature extraction outside KMeans class.
* Use multiple models.
* Add /api/v1/ml_models endpoint.
* Remove Dimension::getID()
* Use just 1 model and fix tests.
* Add detection logic based on rrdr.
* Remove config options related to anomaly detection.
* Make anomaly detection queries configurable.
* Fix ad query duration option.
* Finalize queries in all code paths.
* Check if query was initialized before finalizing it
* Do not leak OWA
* Profile plugin.
* Revert "Profile plugin."
This reverts commit 5c77145d0df7e091d030476c480ab8d9cbceb89e.
* Change context from anomaly_detection to detector_events.
|
|
(#13707)
|
|
* rrdfamily
* rrddim
* rrdset plugin and module names
* rrdset units
* rrdset type
* rrdset family
* rrdset title
* rrdset title more
* rrdset context
* rrdcalctemplate context and removal of context hash from rrdset
* strings statistics
* rrdset name
* rearranged members of rrdset
* eliminate rrdset name hash; rrdcalc chart converted to STRING
* rrdset id, eliminated rrdset hash
* rrdcalc, alarm_entry, alert_config and some of rrdcalctemplate
* rrdcalctemplate
* rrdvar
* eval_variable
* rrddimvar and rrdsetvar
* rrdhost hostname, os and tags
* fix master commits
* added thread cache; implemented string_dup without locks
* faster thread cache
* rrdset and rrddim now use dictionaries for indexing
* rrdhost now uses dictionary
* rrdfamily now uses DICTIONARY
* rrdvar using dictionary instead of AVL
* allocate the right size to rrdvar flag members
* rrdhost remaining char * members to STRING *
* better error handling on indexing
* strings now use a read/write lock to allow parallel searches to the index
* removed AVL support from dictionaries; implemented STRING with native Judy calls
* string releases should be negative
* only 31 bits are allowed for enum flags
* proper locking on strings
* string threading unittest and fixes
* fix lgtm finding
* fixed naming
* stream chart/dimension definitions at the beginning of a streaming session
* thread stack variable is undefined on thread cancel
* rrdcontext garbage collect per host on startup
* worker control in garbage collection
* relaxed deletion of rrdmetrics
* type checking on dictfe
* netdata chart to monitor rrdcontext triggers
* Group chart label updates
* rrdcontext better handling of collected rrdsets
* rrdpush incremental transmition of definitions should use as much buffer as possible
* require 1MB per chart
* empty the sender buffer before enabling metrics streaming
* fill up to 50% of buffer
* reset signaling metrics sending
* use the shared variable for status
* use separate host flag for enabling streaming of metrics
* make sure the flag is clear
* add logging for streaming
* add logging for streaming on buffer overflow
* circular_buffer proper sizing
* removed obsolete logs
* do not execute worker jobs if not necessary
* better messages about compression disabling
* proper use of flags and updating rrdset last access time every time the obsoletion flag is flipped
* monitor stream sender used buffer ratio
* Update exporting unit tests
* no need to compare label value with strcmp
* streaming send workers now monitor bandwidth
* workers now use strings
* streaming receiver monitors incoming bandwidth
* parser shift of worker ids
* minor fixes
* Group chart label updates
* Populate context with dimensions that have data
* Fix chart id
* better shift of parser worker ids
* fix for streaming compression
* properly count received bytes
* ensure LZ4 compression ring buffer does not wrap prematurely
* do not stream empty charts; do not process empty instances in rrdcontext
* need_to_send_chart_definition() does not need an rrdset lock any more
* rrdcontext objects are collected, after data have been written to the db
* better logging of RRDCONTEXT transitions
* always set all variables needed by the worker utilization charts
* implemented double linked list for most objects; eliminated alarm indexes from rrdhost; and many more fixes
* lockless strings design - string_dup() and string_freez() are totally lockless when they dont need to touch Judy - only Judy is protected with a read/write lock
* STRING code re-organization for clarity
* thread_cache improvements; double numbers precision on worker threads
* STRING_ENTRY now shadown STRING, so no duplicate definition is required; string_length() renamed to string_strlen() to follow the paradigm of all other functions, STRING internal statistics are now only compiled with NETDATA_INTERNAL_CHECKS
* rrdhost index by hostname now cleans up; aclk queries of archieved hosts do not index hosts
* Add index to speed up database context searches
* Removed last_updated optimization (was also buggy after latest merge with master)
Co-authored-by: Stelios Fragkakis <52996999+stelfrag@users.noreply.github.com>
Co-authored-by: Vladimir Kobal <vlad@prokk.net>
|
|
* Tier part 1
* Tier part 2
* Tier part 3
* Tier part 4
* Tier part 5
* Fix some ML compilation errors
* fix more conflicts
* pass proper tier
* move metric_uuid from state to RRDDIM
* move aclk_live_status from state to RRDDIM
* move ml_dimension from state to RRDDIM
* abstracted the data collection interface
* support flushing for mem db too
* abstracted the query api
* abstracted latest/oldest time per metric
* cleanup
* store_metric for tier1
* fix for store_metric
* allow multiple tiers, more than 2
* state to tier
* Change storage type in db. Query param to request min, max, sum or average
* Store tier data correctly
* Fix skipping tier page type
* Add tier grouping in the tier
* Fix to handle archived charts (part 1)
* Temp fix for query granularity when requesting tier1 data
* Fix parameters in the correct order and calculate the anomaly based on the anomaly count
* Proper tiering grouping
* Anomaly calculation based on anomaly count
* force type checking on storage handles
* update cmocka tests
* fully dynamic number of storage tiers
* fix static allocation
* configure grouping for all tiers; disable tiers for unittest; disable statsd configuration for private charts mode
* use default page dt using the tiering info
* automatic selection of tier
* fix for automatic selection of tier
* working prototype of dynamic tier selection
* automatic selection of tier done right (I hope)
* ask for the proper tier value, based on the grouping function
* fixes for unittests and load_metric_next()
* fixes for lgtm findings
* minor renames
* add dbengine to page cache size setting
* add dbengine to page cache with malloc
* query engine optimized to loop as little are required based on the view_update_every
* query engine grouping methods now do not assume a constant number of points per group and they allocate memory with OWA
* report db points per tier in jsonwrap
* query planer that switches database tiers on the fly to satisfy the query for the entire timeframe
* dbegnine statistics and documentation (in progress)
* calculate average point duration in db
* handle single point pages the best we can
* handle single point pages even better
* Keep page type in the rrdeng_page_descr
* updated doc
* handle future backwards compatibility - improved statistics
* support &tier=X in queries
* enfore increasing iterations on tiers
* tier 1 is always 1 iteration
* backfilling higher tiers on first data collection
* reversed anomaly bit
* set up to 5 tiers
* natural points should only be offered on tier 0, except a specific tier is selected
* do not allow more than 65535 points of tier0 to be aggregated on any tier
* Work only on actually activated tiers
* fix query interpolation
* fix query interpolation again
* fix lgtm finding
* Activate one tier for now
* backfilling of higher tiers using raw metrics from lower tiers
* fix for crash on start when storage tiers is increased from the default
* more statistics on exit
* fix bug that prevented higher tiers to get any values; added backfilling options
* fixed the statistics log line
* removed limit of 255 iterations per tier; moved the code of freezing rd->tiers[x]->db_metric_handle
* fixed division by zero on zero points_wanted
* removed dead code
* Decide on the descr->type for the type of metric
* dont store metrics on unknown page types
* free db_metric_handle on sql based context queries
* Disable STORAGE_POINT value check in the exporting engine unit tests
* fix for db modes other than dbengine
* fix for aclk archived chart queries destroying db_metric_handles of valid rrddims
* fix left-over freez() instead of OWA freez on median queries
Co-authored-by: Costa Tsaousis <costa@netdata.cloud>
Co-authored-by: Vladimir Kobal <vlad@prokk.net>
|
|
* netdata doubles
* fix cmocka test
* fix cmocka test again
* fix left-overs of long double to NETDATA_DOUBLE
* RRDDIM detached from disk representation; db settings in [db] section of netdata.conf
* update the memory before saving
* rrdset is now detached from file structures too
* on memory mode map, update the memory mapped structures on every iteration
* allow RRD_ID_LENGTH_MAX to be changed
* granularity secs, back to update every
* fix formatting
* more formatting
|
|
Initialization of ML-related structures and threads should happen
when the underlying RRD objects have been fully initialized. Destruction
should happen in the opposite way, ie. before deleting an RRD host/dimension.
|
|
* Move CPU usage stats under netdata charts
Use the hostname in each chart's name, and the machine GUID in each
chart's id.
* Move anomaly_detection.* charts to child host instance.
* Add option to enable/disable streaming of ML-related charts.
* Update priority of prediction/training charts.
|
|
* Increase training thread's max sleep time.
With this change we will only cap the allotted time when it is more than
ten seconds. The previous limit was one second, which had the effect of
scheduling dimensions near the beggining of each training window. This
was not desirable because it would cause high CPU usage on parents with
many children.
* Only exclude netdata.* charts from training.
* Use heartbeat in detection thread.
* Track rusage of prediction thread.
* Track rusage of training thread.
* Add support for random sampling of extracted features.
* Rebase
* Skip RNG when ML is disabled and fix undef behaviour
|
|
* Track anomaly rates with DBEngine.
This commit adds support for tracking anomaly rates with DBEngine. We
do so by creating a single chart with id "anomaly_detection.anomaly_rates" for
each trainable/predictable host, which is responsible for tracking the anomaly
rate of each dimension that we train/predict for that host.
The rrdset->state->is_ar_chart boolean flag is set to true only for anomaly
rates charts. We use this flag to:
- Disable exposing the anomaly rates charts through the functionality
in backends/, exporting/ and streaming/.
- Skip generation of configuration options for the name, algorithm,
multiplier, divisor of each dimension in an anomaly rates chart.
- Skip the creation of health variables for anomaly rates dimensions.
- Skip the chart/dim queue of ACLK.
- Post-process the RRDR result of an anomaly rates chart, so that we can
return a sorted, trimmed number of anomalous dimensions.
In a child/parent configuration where both the child and the parent run
ML for the child, we want to be able to stream the rest of the ML-related
charts to the parent. To be able to do this without any chart name collisions,
the charts are now created on localhost and their IDs and titles have the node's
machine_guid and hostname as a suffix, respectively.
* Fix exporting_engine tests.
* Restore default ML configuration.
The reverted changes where meant for local testing only. This commit
restores the default values that we want to have when someone runs
anomaly detection on their node.
* Set context for anomaly_detection.* charts.
* Check for anomaly rates chart only with a valid pointer.
* Remove duplicate code.
* Use a more descriptive name for id/title pair variable
|
|
* Provide runtime ml info from a new endpoint.
* Add hosts & charts to skip from ML in the /info endpoint.
This information belongs in /info, and not in /ml-info, because the
value of these variables can not change at the agent's runtime.
* Use strdupz instead of strdup.
|
|
* 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 start |