blob: f34423a27affcb2abcfbf148e16b79ea0c26258e [file] [log] [blame]
/*
* Copyright (C) 2017 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#define DEBUG false // STOPSHIP if true
#include "Log.h"
#include "ValueMetricProducer.h"
#include "../guardrail/StatsdStats.h"
#include "../stats_log_util.h"
#include <limits.h>
#include <stdlib.h>
using android::util::FIELD_COUNT_REPEATED;
using android::util::FIELD_TYPE_BOOL;
using android::util::FIELD_TYPE_DOUBLE;
using android::util::FIELD_TYPE_INT32;
using android::util::FIELD_TYPE_INT64;
using android::util::FIELD_TYPE_MESSAGE;
using android::util::FIELD_TYPE_STRING;
using android::util::ProtoOutputStream;
using std::map;
using std::shared_ptr;
using std::unordered_map;
namespace android {
namespace os {
namespace statsd {
// for StatsLogReport
const int FIELD_ID_ID = 1;
const int FIELD_ID_VALUE_METRICS = 7;
const int FIELD_ID_TIME_BASE = 9;
const int FIELD_ID_BUCKET_SIZE = 10;
const int FIELD_ID_DIMENSION_PATH_IN_WHAT = 11;
const int FIELD_ID_IS_ACTIVE = 14;
// for ValueMetricDataWrapper
const int FIELD_ID_DATA = 1;
const int FIELD_ID_SKIPPED = 2;
// for SkippedBuckets
const int FIELD_ID_SKIPPED_START_MILLIS = 3;
const int FIELD_ID_SKIPPED_END_MILLIS = 4;
const int FIELD_ID_SKIPPED_DROP_EVENT = 5;
// for DumpEvent Proto
const int FIELD_ID_BUCKET_DROP_REASON = 1;
const int FIELD_ID_DROP_TIME = 2;
// for ValueMetricData
const int FIELD_ID_DIMENSION_IN_WHAT = 1;
const int FIELD_ID_BUCKET_INFO = 3;
const int FIELD_ID_DIMENSION_LEAF_IN_WHAT = 4;
const int FIELD_ID_SLICE_BY_STATE = 6;
// for ValueBucketInfo
const int FIELD_ID_VALUE_INDEX = 1;
const int FIELD_ID_VALUE_LONG = 2;
const int FIELD_ID_VALUE_DOUBLE = 3;
const int FIELD_ID_VALUES = 9;
const int FIELD_ID_BUCKET_NUM = 4;
const int FIELD_ID_START_BUCKET_ELAPSED_MILLIS = 5;
const int FIELD_ID_END_BUCKET_ELAPSED_MILLIS = 6;
const int FIELD_ID_CONDITION_TRUE_NS = 10;
const Value ZERO_LONG((int64_t)0);
const Value ZERO_DOUBLE((int64_t)0);
// ValueMetric has a minimum bucket size of 10min so that we don't pull too frequently
ValueMetricProducer::ValueMetricProducer(
const ConfigKey& key, const ValueMetric& metric, const int conditionIndex,
const sp<ConditionWizard>& conditionWizard, const int whatMatcherIndex,
const sp<EventMatcherWizard>& matcherWizard, const int pullTagId, const int64_t timeBaseNs,
const int64_t startTimeNs, const sp<StatsPullerManager>& pullerManager,
const unordered_map<int, shared_ptr<Activation>>& eventActivationMap,
const unordered_map<int, vector<shared_ptr<Activation>>>& eventDeactivationMap,
const vector<int>& slicedStateAtoms,
const unordered_map<int, unordered_map<int, int64_t>>& stateGroupMap)
: MetricProducer(metric.id(), key, timeBaseNs, conditionIndex, conditionWizard,
eventActivationMap, eventDeactivationMap, slicedStateAtoms, stateGroupMap),
mWhatMatcherIndex(whatMatcherIndex),
mEventMatcherWizard(matcherWizard),
mPullerManager(pullerManager),
mPullTagId(pullTagId),
mIsPulled(pullTagId != -1),
mMinBucketSizeNs(metric.min_bucket_size_nanos()),
mDimensionSoftLimit(StatsdStats::kAtomDimensionKeySizeLimitMap.find(pullTagId) !=
StatsdStats::kAtomDimensionKeySizeLimitMap.end()
? StatsdStats::kAtomDimensionKeySizeLimitMap.at(pullTagId).first
: StatsdStats::kDimensionKeySizeSoftLimit),
mDimensionHardLimit(StatsdStats::kAtomDimensionKeySizeLimitMap.find(pullTagId) !=
StatsdStats::kAtomDimensionKeySizeLimitMap.end()
? StatsdStats::kAtomDimensionKeySizeLimitMap.at(pullTagId).second
: StatsdStats::kDimensionKeySizeHardLimit),
mUseAbsoluteValueOnReset(metric.use_absolute_value_on_reset()),
mAggregationType(metric.aggregation_type()),
mUseDiff(metric.has_use_diff() ? metric.use_diff() : (mIsPulled ? true : false)),
mValueDirection(metric.value_direction()),
mSkipZeroDiffOutput(metric.skip_zero_diff_output()),
mUseZeroDefaultBase(metric.use_zero_default_base()),
mHasGlobalBase(false),
mCurrentBucketIsInvalid(false),
mMaxPullDelayNs(metric.max_pull_delay_sec() > 0 ? metric.max_pull_delay_sec() * NS_PER_SEC
: StatsdStats::kPullMaxDelayNs),
mSplitBucketForAppUpgrade(metric.split_bucket_for_app_upgrade()),
// Condition timer will be set later within the constructor after pulling events
mConditionTimer(false, timeBaseNs) {
int64_t bucketSizeMills = 0;
if (metric.has_bucket()) {
bucketSizeMills = TimeUnitToBucketSizeInMillisGuardrailed(key.GetUid(), metric.bucket());
} else {
bucketSizeMills = TimeUnitToBucketSizeInMillis(ONE_HOUR);
}
mBucketSizeNs = bucketSizeMills * 1000000;
translateFieldMatcher(metric.value_field(), &mFieldMatchers);
if (metric.has_dimensions_in_what()) {
translateFieldMatcher(metric.dimensions_in_what(), &mDimensionsInWhat);
mContainANYPositionInDimensionsInWhat = HasPositionANY(metric.dimensions_in_what());
mSliceByPositionALL = HasPositionALL(metric.dimensions_in_what());
}
if (metric.links().size() > 0) {
for (const auto& link : metric.links()) {
Metric2Condition mc;
mc.conditionId = link.condition();
translateFieldMatcher(link.fields_in_what(), &mc.metricFields);
translateFieldMatcher(link.fields_in_condition(), &mc.conditionFields);
mMetric2ConditionLinks.push_back(mc);
}
mConditionSliced = true;
}
for (const auto& stateLink : metric.state_link()) {
Metric2State ms;
ms.stateAtomId = stateLink.state_atom_id();
translateFieldMatcher(stateLink.fields_in_what(), &ms.metricFields);
translateFieldMatcher(stateLink.fields_in_state(), &ms.stateFields);
mMetric2StateLinks.push_back(ms);
}
int64_t numBucketsForward = calcBucketsForwardCount(startTimeNs);
mCurrentBucketNum += numBucketsForward;
flushIfNeededLocked(startTimeNs);
if (mIsPulled) {
mPullerManager->RegisterReceiver(mPullTagId, mConfigKey, this, getCurrentBucketEndTimeNs(),
mBucketSizeNs);
}
// Only do this for partial buckets like first bucket. All other buckets should use
// flushIfNeeded to adjust start and end to bucket boundaries.
// Adjust start for partial bucket
mCurrentBucketStartTimeNs = startTimeNs;
mConditionTimer.newBucketStart(mCurrentBucketStartTimeNs);
// Now that activations are processed, start the condition timer if needed.
mConditionTimer.onConditionChanged(mIsActive && mCondition == ConditionState::kTrue,
mCurrentBucketStartTimeNs);
VLOG("value metric %lld created. bucket size %lld start_time: %lld", (long long)metric.id(),
(long long)mBucketSizeNs, (long long)mTimeBaseNs);
}
ValueMetricProducer::~ValueMetricProducer() {
VLOG("~ValueMetricProducer() called");
if (mIsPulled) {
mPullerManager->UnRegisterReceiver(mPullTagId, mConfigKey, this);
}
}
void ValueMetricProducer::onStateChanged(int64_t eventTimeNs, int32_t atomId,
const HashableDimensionKey& primaryKey, int oldState,
int newState) {
VLOG("ValueMetric %lld onStateChanged time %lld, State %d, key %s, %d -> %d",
(long long)mMetricId, (long long)eventTimeNs, atomId, primaryKey.toString().c_str(),
oldState, newState);
// If condition is not true, we do not need to pull for this state change.
if (mCondition != ConditionState::kTrue) {
return;
}
bool isEventLate = eventTimeNs < mCurrentBucketStartTimeNs;
if (isEventLate) {
VLOG("Skip event due to late arrival: %lld vs %lld", (long long)eventTimeNs,
(long long)mCurrentBucketStartTimeNs);
invalidateCurrentBucket(eventTimeNs, BucketDropReason::EVENT_IN_WRONG_BUCKET);
return;
}
mStateChangePrimaryKey.first = atomId;
mStateChangePrimaryKey.second = primaryKey;
if (mIsPulled) {
pullAndMatchEventsLocked(eventTimeNs);
}
mStateChangePrimaryKey.first = 0;
mStateChangePrimaryKey.second = DEFAULT_DIMENSION_KEY;
flushIfNeededLocked(eventTimeNs);
}
void ValueMetricProducer::onSlicedConditionMayChangeLocked(bool overallCondition,
const int64_t eventTime) {
VLOG("Metric %lld onSlicedConditionMayChange", (long long)mMetricId);
}
void ValueMetricProducer::dropDataLocked(const int64_t dropTimeNs) {
StatsdStats::getInstance().noteBucketDropped(mMetricId);
// The current partial bucket is not flushed and does not require a pull,
// so the data is still valid.
flushIfNeededLocked(dropTimeNs);
clearPastBucketsLocked(dropTimeNs);
}
void ValueMetricProducer::clearPastBucketsLocked(const int64_t dumpTimeNs) {
mPastBuckets.clear();
mSkippedBuckets.clear();
}
void ValueMetricProducer::onDumpReportLocked(const int64_t dumpTimeNs,
const bool include_current_partial_bucket,
const bool erase_data,
const DumpLatency dumpLatency,
std::set<string> *str_set,
ProtoOutputStream* protoOutput) {
VLOG("metric %lld dump report now...", (long long)mMetricId);
if (include_current_partial_bucket) {
// For pull metrics, we need to do a pull at bucket boundaries. If we do not do that the
// current bucket will have incomplete data and the next will have the wrong snapshot to do
// a diff against. If the condition is false, we are fine since the base data is reset and
// we are not tracking anything.
bool pullNeeded = mIsPulled && mCondition == ConditionState::kTrue;
if (pullNeeded) {
switch (dumpLatency) {
case FAST:
invalidateCurrentBucket(dumpTimeNs, BucketDropReason::DUMP_REPORT_REQUESTED);
break;
case NO_TIME_CONSTRAINTS:
pullAndMatchEventsLocked(dumpTimeNs);
break;
}
}
flushCurrentBucketLocked(dumpTimeNs, dumpTimeNs);
}
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_ID, (long long)mMetricId);
protoOutput->write(FIELD_TYPE_BOOL | FIELD_ID_IS_ACTIVE, isActiveLocked());
if (mPastBuckets.empty() && mSkippedBuckets.empty()) {
return;
}
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_TIME_BASE, (long long)mTimeBaseNs);
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_BUCKET_SIZE, (long long)mBucketSizeNs);
// Fills the dimension path if not slicing by ALL.
if (!mSliceByPositionALL) {
if (!mDimensionsInWhat.empty()) {
uint64_t dimenPathToken =
protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_ID_DIMENSION_PATH_IN_WHAT);
writeDimensionPathToProto(mDimensionsInWhat, protoOutput);
protoOutput->end(dimenPathToken);
}
}
uint64_t protoToken = protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_ID_VALUE_METRICS);
for (const auto& skippedBucket : mSkippedBuckets) {
uint64_t wrapperToken =
protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED | FIELD_ID_SKIPPED);
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_SKIPPED_START_MILLIS,
(long long)(NanoToMillis(skippedBucket.bucketStartTimeNs)));
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_SKIPPED_END_MILLIS,
(long long)(NanoToMillis(skippedBucket.bucketEndTimeNs)));
for (const auto& dropEvent : skippedBucket.dropEvents) {
uint64_t dropEventToken = protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED |
FIELD_ID_SKIPPED_DROP_EVENT);
protoOutput->write(FIELD_TYPE_INT32 | FIELD_ID_BUCKET_DROP_REASON, dropEvent.reason);
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_DROP_TIME,
(long long)(NanoToMillis(dropEvent.dropTimeNs)));
;
protoOutput->end(dropEventToken);
}
protoOutput->end(wrapperToken);
}
for (const auto& pair : mPastBuckets) {
const MetricDimensionKey& dimensionKey = pair.first;
VLOG(" dimension key %s", dimensionKey.toString().c_str());
uint64_t wrapperToken =
protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED | FIELD_ID_DATA);
// First fill dimension.
if (mSliceByPositionALL) {
uint64_t dimensionToken =
protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_ID_DIMENSION_IN_WHAT);
writeDimensionToProto(dimensionKey.getDimensionKeyInWhat(), str_set, protoOutput);
protoOutput->end(dimensionToken);
} else {
writeDimensionLeafNodesToProto(dimensionKey.getDimensionKeyInWhat(),
FIELD_ID_DIMENSION_LEAF_IN_WHAT, str_set, protoOutput);
}
// Then fill slice_by_state.
for (auto state : dimensionKey.getStateValuesKey().getValues()) {
uint64_t stateToken = protoOutput->start(FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED |
FIELD_ID_SLICE_BY_STATE);
writeStateToProto(state, protoOutput);
protoOutput->end(stateToken);
}
// Then fill bucket_info (ValueBucketInfo).
for (const auto& bucket : pair.second) {
uint64_t bucketInfoToken = protoOutput->start(
FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED | FIELD_ID_BUCKET_INFO);
if (bucket.mBucketEndNs - bucket.mBucketStartNs != mBucketSizeNs) {
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_START_BUCKET_ELAPSED_MILLIS,
(long long)NanoToMillis(bucket.mBucketStartNs));
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_END_BUCKET_ELAPSED_MILLIS,
(long long)NanoToMillis(bucket.mBucketEndNs));
} else {
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_BUCKET_NUM,
(long long)(getBucketNumFromEndTimeNs(bucket.mBucketEndNs)));
}
// only write the condition timer value if the metric has a condition.
if (mConditionTrackerIndex >= 0) {
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_CONDITION_TRUE_NS,
(long long)bucket.mConditionTrueNs);
}
for (int i = 0; i < (int)bucket.valueIndex.size(); i++) {
int index = bucket.valueIndex[i];
const Value& value = bucket.values[i];
uint64_t valueToken = protoOutput->start(
FIELD_TYPE_MESSAGE | FIELD_COUNT_REPEATED | FIELD_ID_VALUES);
protoOutput->write(FIELD_TYPE_INT32 | FIELD_ID_VALUE_INDEX,
index);
if (value.getType() == LONG) {
protoOutput->write(FIELD_TYPE_INT64 | FIELD_ID_VALUE_LONG,
(long long)value.long_value);
VLOG("\t bucket [%lld - %lld] value %d: %lld", (long long)bucket.mBucketStartNs,
(long long)bucket.mBucketEndNs, index, (long long)value.long_value);
} else if (value.getType() == DOUBLE) {
protoOutput->write(FIELD_TYPE_DOUBLE | FIELD_ID_VALUE_DOUBLE,
value.double_value);
VLOG("\t bucket [%lld - %lld] value %d: %.2f", (long long)bucket.mBucketStartNs,
(long long)bucket.mBucketEndNs, index, value.double_value);
} else {
VLOG("Wrong value type for ValueMetric output: %d", value.getType());
}
protoOutput->end(valueToken);
}
protoOutput->end(bucketInfoToken);
}
protoOutput->end(wrapperToken);
}
protoOutput->end(protoToken);
VLOG("metric %lld dump report now...", (long long)mMetricId);
if (erase_data) {
mPastBuckets.clear();
mSkippedBuckets.clear();
}
}
void ValueMetricProducer::invalidateCurrentBucketWithoutResetBase(const int64_t dropTimeNs,
const BucketDropReason reason) {
if (!mCurrentBucketIsInvalid) {
// Only report to StatsdStats once per invalid bucket.
StatsdStats::getInstance().noteInvalidatedBucket(mMetricId);
}
if (!maxDropEventsReached()) {
mCurrentSkippedBucket.dropEvents.emplace_back(buildDropEvent(dropTimeNs, reason));
}
mCurrentBucketIsInvalid = true;
}
void ValueMetricProducer::invalidateCurrentBucket(const int64_t dropTimeNs,
const BucketDropReason reason) {
invalidateCurrentBucketWithoutResetBase(dropTimeNs, reason);
resetBase();
}
void ValueMetricProducer::resetBase() {
for (auto& slice : mCurrentBaseInfo) {
for (auto& baseInfo : slice.second) {
baseInfo.hasBase = false;
baseInfo.hasCurrentState = false;
}
}
mHasGlobalBase = false;
}
// Handle active state change. Active state change is treated like a condition change:
// - drop bucket if active state change event arrives too late
// - if condition is true, pull data on active state changes
// - ConditionTimer tracks changes based on AND of condition and active state.
void ValueMetricProducer::onActiveStateChangedLocked(const int64_t& eventTimeNs) {
bool isEventTooLate = eventTimeNs < mCurrentBucketStartTimeNs;
if (isEventTooLate) {
// Drop bucket because event arrived too late, ie. we are missing data for this bucket.
StatsdStats::getInstance().noteLateLogEventSkipped(mMetricId);
invalidateCurrentBucket(eventTimeNs, BucketDropReason::EVENT_IN_WRONG_BUCKET);
}
// Call parent method once we've verified the validity of current bucket.
MetricProducer::onActiveStateChangedLocked(eventTimeNs);
if (ConditionState::kTrue != mCondition) {
return;
}
// Pull on active state changes.
if (!isEventTooLate) {
if (mIsPulled) {
pullAndMatchEventsLocked(eventTimeNs);
}
// When active state changes from true to false, clear diff base but don't
// reset other counters as we may accumulate more value in the bucket.
if (mUseDiff && !mIsActive) {
resetBase();
}
}
flushIfNeededLocked(eventTimeNs);
// Let condition timer know of new active state.
mConditionTimer.onConditionChanged(mIsActive, eventTimeNs);
}
void ValueMetricProducer::onConditionChangedLocked(const bool condition,
const int64_t eventTimeNs) {
ConditionState newCondition = condition ? ConditionState::kTrue : ConditionState::kFalse;
bool isEventTooLate = eventTimeNs < mCurrentBucketStartTimeNs;
// If the config is not active, skip the event.
if (!mIsActive) {
mCondition = isEventTooLate ? ConditionState::kUnknown : newCondition;
return;
}
// If the event arrived late, mark the bucket as invalid and skip the event.
if (isEventTooLate) {
VLOG("Skip event due to late arrival: %lld vs %lld", (long long)eventTimeNs,
(long long)mCurrentBucketStartTimeNs);
StatsdStats::getInstance().noteLateLogEventSkipped(mMetricId);
StatsdStats::getInstance().noteConditionChangeInNextBucket(mMetricId);
invalidateCurrentBucket(eventTimeNs, BucketDropReason::EVENT_IN_WRONG_BUCKET);
mCondition = ConditionState::kUnknown;
mConditionTimer.onConditionChanged(mCondition, eventTimeNs);
return;
}
// If the previous condition was unknown, mark the bucket as invalid
// because the bucket will contain partial data. For example, the condition
// change might happen close to the end of the bucket and we might miss a
// lot of data.
//
// We still want to pull to set the base.
if (mCondition == ConditionState::kUnknown) {
invalidateCurrentBucket(eventTimeNs, BucketDropReason::CONDITION_UNKNOWN);
}
// Pull and match for the following condition change cases:
// unknown/false -> true - condition changed
// true -> false - condition changed
// true -> true - old condition was true so we can flush the bucket at the
// end if needed.
//
// We don’t need to pull for unknown -> false or false -> false.
//
// onConditionChangedLocked might happen on bucket boundaries if this is
// called before #onDataPulled.
if (mIsPulled &&
(newCondition == ConditionState::kTrue || mCondition == ConditionState::kTrue)) {
pullAndMatchEventsLocked(eventTimeNs);
}
// For metrics that use diff, when condition changes from true to false,
// clear diff base but don't reset other counts because we may accumulate
// more value in the bucket.
if (mUseDiff &&
(mCondition == ConditionState::kTrue && newCondition == ConditionState::kFalse)) {
resetBase();
}
// Update condition state after pulling.
mCondition = newCondition;
flushIfNeededLocked(eventTimeNs);
mConditionTimer.onConditionChanged(mCondition, eventTimeNs);
}
void ValueMetricProducer::prepareFirstBucketLocked() {
// Kicks off the puller immediately if condition is true and diff based.
if (mIsActive && mIsPulled && mCondition == ConditionState::kTrue && mUseDiff) {
pullAndMatchEventsLocked(mCurrentBucketStartTimeNs);
}
}
void ValueMetricProducer::pullAndMatchEventsLocked(const int64_t timestampNs) {
vector<std::shared_ptr<LogEvent>> allData;
if (!mPullerManager->Pull(mPullTagId, mConfigKey, &allData)) {
ALOGE("Stats puller failed for tag: %d at %lld", mPullTagId, (long long)timestampNs);
invalidateCurrentBucket(timestampNs, BucketDropReason::PULL_FAILED);
return;
}
accumulateEvents(allData, timestampNs, timestampNs);
}
int64_t ValueMetricProducer::calcPreviousBucketEndTime(const int64_t currentTimeNs) {
return mTimeBaseNs + ((currentTimeNs - mTimeBaseNs) / mBucketSizeNs) * mBucketSizeNs;
}
// By design, statsd pulls data at bucket boundaries using AlarmManager. These pulls are likely
// to be delayed. Other events like condition changes or app upgrade which are not based on
// AlarmManager might have arrived earlier and close the bucket.
void ValueMetricProducer::onDataPulled(const std::vector<std::shared_ptr<LogEvent>>& allData,
bool pullSuccess, int64_t originalPullTimeNs) {
std::lock_guard<std::mutex> lock(mMutex);
if (mCondition == ConditionState::kTrue) {
// If the pull failed, we won't be able to compute a diff.
if (!pullSuccess) {
invalidateCurrentBucket(originalPullTimeNs, BucketDropReason::PULL_FAILED);
} else {
bool isEventLate = originalPullTimeNs < getCurrentBucketEndTimeNs();
if (isEventLate) {
// If the event is late, we are in the middle of a bucket. Just
// process the data without trying to snap the data to the nearest bucket.
accumulateEvents(allData, originalPullTimeNs, originalPullTimeNs);
} else {
// For scheduled pulled data, the effective event time is snap to the nearest
// bucket end. In the case of waking up from a deep sleep state, we will
// attribute to the previous bucket end. If the sleep was long but not very
// long, we will be in the immediate next bucket. Previous bucket may get a
// larger number as we pull at a later time than real bucket end.
//
// If the sleep was very long, we skip more than one bucket before sleep. In
// this case, if the diff base will be cleared and this new data will serve as
// new diff base.
int64_t bucketEndTime = calcPreviousBucketEndTime(originalPullTimeNs) - 1;
StatsdStats::getInstance().noteBucketBoundaryDelayNs(
mMetricId, originalPullTimeNs - bucketEndTime);
accumulateEvents(allData, originalPullTimeNs, bucketEndTime);
}
}
}
// We can probably flush the bucket. Since we used bucketEndTime when calling
// #onMatchedLogEventInternalLocked, the current bucket will not have been flushed.
flushIfNeededLocked(originalPullTimeNs);
}
void ValueMetricProducer::accumulateEvents(const std::vector<std::shared_ptr<LogEvent>>& allData,
int64_t originalPullTimeNs, int64_t eventElapsedTimeNs) {
bool isEventLate = eventElapsedTimeNs < mCurrentBucketStartTimeNs;
if (isEventLate) {
VLOG("Skip bucket end pull due to late arrival: %lld vs %lld",
(long long)eventElapsedTimeNs, (long long)mCurrentBucketStartTimeNs);
StatsdStats::getInstance().noteLateLogEventSkipped(mMetricId);
invalidateCurrentBucket(eventElapsedTimeNs, BucketDropReason::EVENT_IN_WRONG_BUCKET);
return;
}
const int64_t elapsedRealtimeNs = getElapsedRealtimeNs();
const int64_t pullDelayNs = elapsedRealtimeNs - originalPullTimeNs;
StatsdStats::getInstance().notePullDelay(mPullTagId, pullDelayNs);
if (pullDelayNs > mMaxPullDelayNs) {
ALOGE("Pull finish too late for atom %d, longer than %lld", mPullTagId,
(long long)mMaxPullDelayNs);
StatsdStats::getInstance().notePullExceedMaxDelay(mPullTagId);
// We are missing one pull from the bucket which means we will not have a complete view of
// what's going on.
invalidateCurrentBucket(eventElapsedTimeNs, BucketDropReason::PULL_DELAYED);
return;
}
if (allData.size() == 0) {
VLOG("Data pulled is empty");
StatsdStats::getInstance().noteEmptyData(mPullTagId);
}
mMatchedMetricDimensionKeys.clear();
for (const auto& data : allData) {
LogEvent localCopy = data->makeCopy();
if (mEventMatcherWizard->matchLogEvent(localCopy, mWhatMatcherIndex) ==
MatchingState::kMatched) {
localCopy.setElapsedTimestampNs(eventElapsedTimeNs);
onMatchedLogEventLocked(mWhatMatcherIndex, localCopy);
}
}
// If a key that is:
// 1. Tracked in mCurrentSlicedBucket and
// 2. A superset of the current mStateChangePrimaryKey
// was not found in the new pulled data (i.e. not in mMatchedDimensionInWhatKeys)
// then we need to reset the base.
for (auto& slice : mCurrentSlicedBucket) {
const auto& whatKey = slice.first.getDimensionKeyInWhat();
bool presentInPulledData =
mMatchedMetricDimensionKeys.find(whatKey) != mMatchedMetricDimensionKeys.end();
if (!presentInPulledData && whatKey.contains(mStateChangePrimaryKey.second)) {
auto it = mCurrentBaseInfo.find(whatKey);
for (auto& baseInfo : it->second) {
baseInfo.hasBase = false;
baseInfo.hasCurrentState = false;
}
}
}
mMatchedMetricDimensionKeys.clear();
mHasGlobalBase = true;
// If we reach the guardrail, we might have dropped some data which means the bucket is
// incomplete.
//
// The base also needs to be reset. If we do not have the full data, we might
// incorrectly compute the diff when mUseZeroDefaultBase is true since an existing key
// might be missing from mCurrentSlicedBucket.
if (hasReachedGuardRailLimit()) {
invalidateCurrentBucket(eventElapsedTimeNs, BucketDropReason::DIMENSION_GUARDRAIL_REACHED);
mCurrentSlicedBucket.clear();
}
}
void ValueMetricProducer::dumpStatesLocked(FILE* out, bool verbose) const {
if (mCurrentSlicedBucket.size() == 0) {
return;
}
fprintf(out, "ValueMetric %lld dimension size %lu\n", (long long)mMetricId,
(unsigned long)mCurrentSlicedBucket.size());
if (verbose) {
for (const auto& it : mCurrentSlicedBucket) {
for (const auto& interval : it.second) {
fprintf(out, "\t(what)%s\t(states)%s (value)%s\n",
it.first.getDimensionKeyInWhat().toString().c_str(),
it.first.getStateValuesKey().toString().c_str(),
interval.value.toString().c_str());
}
}
}
}
bool ValueMetricProducer::hasReachedGuardRailLimit() const {
return mCurrentSlicedBucket.size() >= mDimensionHardLimit;
}
bool ValueMetricProducer::hitGuardRailLocked(const MetricDimensionKey& newKey) {
// ===========GuardRail==============
// 1. Report the tuple count if the tuple count > soft limit
if (mCurrentSlicedBucket.find(newKey) != mCurrentSlicedBucket.end()) {
return false;
}
if (mCurrentSlicedBucket.size() > mDimensionSoftLimit - 1) {
size_t newTupleCount = mCurrentSlicedBucket.size() + 1;
StatsdStats::getInstance().noteMetricDimensionSize(mConfigKey, mMetricId, newTupleCount);
// 2. Don't add more tuples, we are above the allowed threshold. Drop the data.
if (hasReachedGuardRailLimit()) {
ALOGE("ValueMetric %lld dropping data for dimension key %s", (long long)mMetricId,
newKey.toString().c_str());
StatsdStats::getInstance().noteHardDimensionLimitReached(mMetricId);
return true;
}
}
return false;
}
bool ValueMetricProducer::hitFullBucketGuardRailLocked(const MetricDimensionKey& newKey) {
// ===========GuardRail==============
// 1. Report the tuple count if the tuple count > soft limit
if (mCurrentFullBucket.find(newKey) != mCurrentFullBucket.end()) {
return false;
}
if (mCurrentFullBucket.size() > mDimensionSoftLimit - 1) {
size_t newTupleCount = mCurrentFullBucket.size() + 1;
// 2. Don't add more tuples, we are above the allowed threshold. Drop the data.
if (newTupleCount > mDimensionHardLimit) {
ALOGE("ValueMetric %lld dropping data for full bucket dimension key %s",
(long long)mMetricId,
newKey.toString().c_str());
return true;
}
}
return false;
}
bool getDoubleOrLong(const LogEvent& event, const Matcher& matcher, Value& ret) {
for (const FieldValue& value : event.getValues()) {
if (value.mField.matches(matcher)) {
switch (value.mValue.type) {
case INT:
ret.setLong(value.mValue.int_value);
break;
case LONG:
ret.setLong(value.mValue.long_value);
break;
case FLOAT:
ret.setDouble(value.mValue.float_value);
break;
case DOUBLE:
ret.setDouble(value.mValue.double_value);
break;
default:
return false;
break;
}
return true;
}
}
return false;
}
void ValueMetricProducer::onMatchedLogEventInternalLocked(
const size_t matcherIndex, const MetricDimensionKey& eventKey,
const ConditionKey& conditionKey, bool condition, const LogEvent& event,
const map<int, HashableDimensionKey>& statePrimaryKeys) {
auto whatKey = eventKey.getDimensionKeyInWhat();
auto stateKey = eventKey.getStateValuesKey();
// Skip this event if a state changed occurred for a different primary key.
auto it = statePrimaryKeys.find(mStateChangePrimaryKey.first);
// Check that both the atom id and the primary key are equal.
if (it != statePrimaryKeys.end() && it->second != mStateChangePrimaryKey.second) {
VLOG("ValueMetric skip event with primary key %s because state change primary key "
"is %s",
it->second.toString().c_str(), mStateChangePrimaryKey.second.toString().c_str());
return;
}
int64_t eventTimeNs = event.GetElapsedTimestampNs();
if (eventTimeNs < mCurrentBucketStartTimeNs) {
VLOG("Skip event due to late arrival: %lld vs %lld", (long long)eventTimeNs,
(long long)mCurrentBucketStartTimeNs);
return;
}
mMatchedMetricDimensionKeys.insert(whatKey);
if (!mIsPulled) {
// We cannot flush without doing a pull first.
flushIfNeededLocked(eventTimeNs);
}
// We should not accumulate the data for pushed metrics when the condition is false.
bool shouldSkipForPushMetric = !mIsPulled && !condition;
// For pulled metrics, there are two cases:
// - to compute diffs, we need to process all the state changes
// - for non-diffs metrics, we should ignore the data if the condition wasn't true. If we have a
// state change from
// + True -> True: we should process the data, it might be a bucket boundary
// + True -> False: we als need to process the data.
bool shouldSkipForPulledMetric = mIsPulled && !mUseDiff
&& mCondition != ConditionState::kTrue;
if (shouldSkipForPushMetric || shouldSkipForPulledMetric) {
VLOG("ValueMetric skip event because condition is false");
return;
}
if (hitGuardRailLocked(eventKey)) {
return;
}
vector<BaseInfo>& baseInfos = mCurrentBaseInfo[whatKey];
if (baseInfos.size() < mFieldMatchers.size()) {
VLOG("Resizing number of intervals to %d", (int)mFieldMatchers.size());
baseInfos.resize(mFieldMatchers.size());
}
for (auto baseInfo : baseInfos) {
if (!baseInfo.hasCurrentState) {
baseInfo.currentState = DEFAULT_DIMENSION_KEY;
baseInfo.hasCurrentState = true;
}
}
// We need to get the intervals stored with the previous state key so we can
// close these value intervals.
const auto oldStateKey = baseInfos[0].currentState;
vector<Interval>& intervals = mCurrentSlicedBucket[MetricDimensionKey(whatKey, oldStateKey)];
if (intervals.size() < mFieldMatchers.size()) {
VLOG("Resizing number of intervals to %d", (int)mFieldMatchers.size());
intervals.resize(mFieldMatchers.size());
}
// We only use anomaly detection under certain cases.
// N.B.: The anomaly detection cases were modified in order to fix an issue with value metrics
// containing multiple values. We tried to retain all previous behaviour, but we are unsure the
// previous behaviour was correct. At the time of the fix, anomaly detection had no owner.
// Whoever next works on it should look into the cases where it is triggered in this function.
// Discussion here: http://ag/6124370.
bool useAnomalyDetection = true;
for (int i = 0; i < (int)mFieldMatchers.size(); i++) {
const Matcher& matcher = mFieldMatchers[i];
BaseInfo& baseInfo = baseInfos[i];
Interval& interval = intervals[i];
interval.valueIndex = i;
Value value;
if (!getDoubleOrLong(event, matcher, value)) {
VLOG("Failed to get value %d from event %s", i, event.ToString().c_str());
StatsdStats::getInstance().noteBadValueType(mMetricId);
return;
}
interval.seenNewData = true;
if (mUseDiff) {
if (!baseInfo.hasBase) {
if (mHasGlobalBase && mUseZeroDefaultBase) {
// The bucket has global base. This key does not.
// Optionally use zero as base.
baseInfo.base = (value.type == LONG ? ZERO_LONG : ZERO_DOUBLE);
baseInfo.hasBase = true;
} else {
// no base. just update base and return.
baseInfo.base = value;
baseInfo.hasBase = true;
// If we're missing a base, do not use anomaly detection on incomplete data
useAnomalyDetection = false;
// Continue (instead of return) here in order to set baseInfo.base and
// baseInfo.hasBase for other baseInfos
continue;
}
}
Value diff;
switch (mValueDirection) {
case ValueMetric::INCREASING:
if (value >= baseInfo.base) {
diff = value - baseInfo.base;
} else if (mUseAbsoluteValueOnReset) {
diff = value;
} else {
VLOG("Unexpected decreasing value");
StatsdStats::getInstance().notePullDataError(mPullTagId);
baseInfo.base = value;
// If we've got bad data, do not use anomaly detection
useAnomalyDetection = false;
continue;
}
break;
case ValueMetric::DECREASING:
if (baseInfo.base >= value) {
diff = baseInfo.base - value;
} else if (mUseAbsoluteValueOnReset) {
diff = value;
} else {
VLOG("Unexpected increasing value");
StatsdStats::getInstance().notePullDataError(mPullTagId);
baseInfo.base = value;
// If we've got bad data, do not use anomaly detection
useAnomalyDetection = false;
continue;
}
break;
case ValueMetric::ANY:
diff = value - baseInfo.base;
break;
default:
break;
}
baseInfo.base = value;
value = diff;
}
if (interval.hasValue) {
switch (mAggregationType) {
case ValueMetric::SUM:
// for AVG, we add up and take average when flushing the bucket
case ValueMetric::AVG:
interval.value += value;
break;
case ValueMetric::MIN:
interval.value = std::min(value, interval.value);
break;
case ValueMetric::MAX:
interval.value = std::max(value, interval.value);
break;
default:
break;
}
} else {
interval.value = value;
interval.hasValue = true;
}
interval.sampleSize += 1;
baseInfo.currentState = stateKey;
}
// Only trigger the tracker if all intervals are correct
if (useAnomalyDetection) {
// TODO: propgate proper values down stream when anomaly support doubles
long wholeBucketVal = intervals[0].value.long_value;
auto prev = mCurrentFullBucket.find(eventKey);
if (prev != mCurrentFullBucket.end()) {
wholeBucketVal += prev->second;
}
for (auto& tracker : mAnomalyTrackers) {
tracker->detectAndDeclareAnomaly(eventTimeNs, mCurrentBucketNum, mMetricId, eventKey,
wholeBucketVal);
}
}
}
// For pulled metrics, we always need to make sure we do a pull before flushing the bucket
// if mCondition is true!
void ValueMetricProducer::flushIfNeededLocked(const int64_t& eventTimeNs) {
int64_t currentBucketEndTimeNs = getCurrentBucketEndTimeNs();
if (eventTimeNs < currentBucketEndTimeNs) {
VLOG("eventTime is %lld, less than current bucket end time %lld", (long long)eventTimeNs,
(long long)(currentBucketEndTimeNs));
return;
}
int64_t numBucketsForward = calcBucketsForwardCount(eventTimeNs);
int64_t nextBucketStartTimeNs = currentBucketEndTimeNs + (numBucketsForward - 1) * mBucketSizeNs;
flushCurrentBucketLocked(eventTimeNs, nextBucketStartTimeNs);
}
int64_t ValueMetricProducer::calcBucketsForwardCount(const int64_t& eventTimeNs) const {
int64_t currentBucketEndTimeNs = getCurrentBucketEndTimeNs();
if (eventTimeNs < currentBucketEndTimeNs) {
return 0;
}
return 1 + (eventTimeNs - currentBucketEndTimeNs) / mBucketSizeNs;
}
void ValueMetricProducer::flushCurrentBucketLocked(const int64_t& eventTimeNs,
const int64_t& nextBucketStartTimeNs) {
if (mCondition == ConditionState::kUnknown) {
StatsdStats::getInstance().noteBucketUnknownCondition(mMetricId);
}
int64_t numBucketsForward = calcBucketsForwardCount(eventTimeNs);
if (numBucketsForward > 1) {
VLOG("Skipping forward %lld buckets", (long long)numBucketsForward);
StatsdStats::getInstance().noteSkippedForwardBuckets(mMetricId);
// Something went wrong. Maybe the device was sleeping for a long time. It is better
// to mark the current bucket as invalid. The last pull might have been successful through.
invalidateCurrentBucketWithoutResetBase(eventTimeNs,
BucketDropReason::MULTIPLE_BUCKETS_SKIPPED);
}
VLOG("finalizing bucket for %ld, dumping %d slices", (long)mCurrentBucketStartTimeNs,
(int)mCurrentSlicedBucket.size());
int64_t fullBucketEndTimeNs = getCurrentBucketEndTimeNs();
int64_t bucketEndTime = eventTimeNs < fullBucketEndTimeNs ? eventTimeNs : fullBucketEndTimeNs;
// Close the current bucket.
int64_t conditionTrueDuration = mConditionTimer.newBucketStart(bucketEndTime);
bool isBucketLargeEnough = bucketEndTime - mCurrentBucketStartTimeNs >= mMinBucketSizeNs;
if (!isBucketLargeEnough) {
if (!maxDropEventsReached()) {
mCurrentSkippedBucket.dropEvents.emplace_back(
buildDropEvent(eventTimeNs, BucketDropReason::BUCKET_TOO_SMALL));
}
}
if (isBucketLargeEnough && !mCurrentBucketIsInvalid) {
// The current bucket is large enough to keep.
for (const auto& slice : mCurrentSlicedBucket) {
ValueBucket bucket = buildPartialBucket(bucketEndTime, slice.second);
bucket.mConditionTrueNs = conditionTrueDuration;
// it will auto create new vector of ValuebucketInfo if the key is not found.
if (bucket.valueIndex.size() > 0) {
auto& bucketList = mPastBuckets[slice.first];
bucketList.push_back(bucket);
}
}
} else {
mCurrentSkippedBucket.bucketStartTimeNs = mCurrentBucketStartTimeNs;
mCurrentSkippedBucket.bucketEndTimeNs = bucketEndTime;
mSkippedBuckets.emplace_back(mCurrentSkippedBucket);
}
appendToFullBucket(eventTimeNs, fullBucketEndTimeNs);
initCurrentSlicedBucket(nextBucketStartTimeNs);
// Update the condition timer again, in case we skipped buckets.
mConditionTimer.newBucketStart(nextBucketStartTimeNs);
mCurrentBucketNum += numBucketsForward;
}
ValueBucket ValueMetricProducer::buildPartialBucket(int64_t bucketEndTime,
const std::vector<Interval>& intervals) {
ValueBucket bucket;
bucket.mBucketStartNs = mCurrentBucketStartTimeNs;
bucket.mBucketEndNs = bucketEndTime;
for (const auto& interval : intervals) {
if (interval.hasValue) {
// skip the output if the diff is zero
if (mSkipZeroDiffOutput && mUseDiff && interval.value.isZero()) {
continue;
}
bucket.valueIndex.push_back(interval.valueIndex);
if (mAggregationType != ValueMetric::AVG) {
bucket.values.push_back(interval.value);
} else {
double sum = interval.value.type == LONG ? (double)interval.value.long_value
: interval.value.double_value;
bucket.values.push_back(Value((double)sum / interval.sampleSize));
}
}
}
return bucket;
}
void ValueMetricProducer::initCurrentSlicedBucket(int64_t nextBucketStartTimeNs) {
StatsdStats::getInstance().noteBucketCount(mMetricId);
// Cleanup data structure to aggregate values.
for (auto it = mCurrentSlicedBucket.begin(); it != mCurrentSlicedBucket.end();) {
bool obsolete = true;
for (auto& interval : it->second) {
interval.hasValue = false;
interval.sampleSize = 0;
if (interval.seenNewData) {
obsolete = false;
}
interval.seenNewData = false;
}
if (obsolete) {
it = mCurrentSlicedBucket.erase(it);
} else {
it++;
}
// TODO: remove mCurrentBaseInfo entries when obsolete
}
mCurrentBucketIsInvalid = false;
mCurrentSkippedBucket.reset();
// If we do not have a global base when the condition is true,
// we will have incomplete bucket for the next bucket.
if (mUseDiff && !mHasGlobalBase && mCondition) {
mCurrentBucketIsInvalid = false;
}
mCurrentBucketStartTimeNs = nextBucketStartTimeNs;
VLOG("metric %lld: new bucket start time: %lld", (long long)mMetricId,
(long long)mCurrentBucketStartTimeNs);
}
void ValueMetricProducer::appendToFullBucket(int64_t eventTimeNs, int64_t fullBucketEndTimeNs) {
bool isFullBucketReached = eventTimeNs > fullBucketEndTimeNs;
if (mCurrentBucketIsInvalid) {
if (isFullBucketReached) {
// If the bucket is invalid, we ignore the full bucket since it contains invalid data.
mCurrentFullBucket.clear();
}
// Current bucket is invalid, we do not add it to the full bucket.
return;
}
if (isFullBucketReached) { // If full bucket, send to anomaly tracker.
// Accumulate partial buckets with current value and then send to anomaly tracker.
if (mCurrentFullBucket.size() > 0) {
for (const auto& slice : mCurrentSlicedBucket) {
if (hitFullBucketGuardRailLocked(slice.first)) {
continue;
}
// TODO: fix this when anomaly can accept double values
auto& interval = slice.second[0];
if (interval.hasValue) {
mCurrentFullBucket[slice.first] += interval.value.long_value;
}
}
for (const auto& slice : mCurrentFullBucket) {
for (auto& tracker : mAnomalyTrackers) {
if (tracker != nullptr) {
tracker->addPastBucket(slice.first, slice.second, mCurrentBucketNum);
}
}
}
mCurrentFullBucket.clear();
} else {
// Skip aggregating the partial buckets since there's no previous partial bucket.
for (const auto& slice : mCurrentSlicedBucket) {
for (auto& tracker : mAnomalyTrackers) {
if (tracker != nullptr) {
// TODO: fix this when anomaly can accept double values
auto& interval = slice.second[0];
if (interval.hasValue) {
tracker->addPastBucket(slice.first, interval.value.long_value,
mCurrentBucketNum);
}
}
}
}
}
} else {
// Accumulate partial bucket.
for (const auto& slice : mCurrentSlicedBucket) {
// TODO: fix this when anomaly can accept double values
auto& interval = slice.second[0];
if (interval.hasValue) {
mCurrentFullBucket[slice.first] += interval.value.long_value;
}
}
}
}
size_t ValueMetricProducer::byteSizeLocked() const {
size_t totalSize = 0;
for (const auto& pair : mPastBuckets) {
totalSize += pair.second.size() * kBucketSize;
}
return totalSize;
}
} // namespace statsd
} // namespace os
} // namespace android