1. Hash the strings in metric dimensions.
2. Optimize the timestamp encoding in bucket.
Use bucket num for full bucket and millis for
partial bucket.
3. Encode the dimension path per metric and avoid
deduping it across dimensons.
Test: statsd test
Change-Id: I18f69654de85edb21a9c835c73edead756295e05
BUG: b/77813755
When statsd reconnects to logd, statsd will read all logs from buffer again. To prevent us from
reprocessing old events, we do the following:
1. At any given moment, record the largest timestamp(T_max) and last timestamp (check point) that
we've seen before.
2. When reconnection happens, we look for the check point until we see a new log with a timestamp
larger than T_max.
-> If we found the CP, resume after the CP. Success
-> If we can't find CP, there is definitely log loss. We reset all configs.
Note:
1. Logd has an API to read logs after a certain timestamp. But this api is vulnerable to
time changes from Settings. So we cannot rely on it.
2. If logd inserts a new log (with older timestamp) before CP, we cannot detect it. It's not
possible to detect it without record all timestamps we have seen.
Test: statsd_test
Bug: 77813113
Change-Id: Ic3fdb47230807606ab11dc994cb162194adb8448
The previous scheme captured periodic snapshots for each config with
complex logic that's unnecessary and wasted memory. We actually don't
need to store any snapshots since we just convert the current state
into a snapshot and also include the deltas (change events) since the
previous report until now.
To make the system more robust, we also include up to 100 of the
deleted apps in the uid map.
Also, fix the wiring of the partial buckets so the metric producers
form partial buckets on both app upgrade and removal, but not on
installation of a new app.
Also, we update StatsCompanionService to also include disabled apps.
Bug: 77607583
Test: Verified unit-tests pass and added new e2e tests.
Change-Id: I98e1f544d6e6571545ae1581c4cebab807596f51
Roughly check the config every hour to see whether the ttl expired.
If so, read the config from disk and recreate the metric manager.
Test: statsd test
BUG: b/77274363
Change-Id: I16838afe5bbe966c3a0f638869751f9b59a5a259
It's tricky to determine the source of the metrics on a device
currently since we can take the union of multiple configs and send
only one giant statsd_config into statsd. We will use the int64 field
to track the sub config id's and the int32 field to track the version
for each sub config, but the fields are named more generically as
annotations.
The annotations are available in both the reports and metadata.
Test: Check that all unit-tests pass on marlin-eng
Bug: 77327261
Change-Id: Ic37c549c8b2991676f69948c515156765c9f5108
Previously tried an optimization that results in corrupted proto
output. This changes to a safer approach of storing the snapshot data
in memory and only converting to proto output when the
ProtoOutputStream is provided.
Also fixes a security issue when trying to invoke triggerUidSnapshot
since we forgot to use SCS' permissions.
Test: Added a unit-test to verify output of StatsLogProcessor.
Bug: 76231867
Change-Id: Id410ce3505fda9d71caa71942ef3068b55872c66
+ Metrics will do flushIfNeeded() to correctly move the clock and informing
AnomalyTracker the past bucket info, and then clear past buckets.
+ We will still keep the current bucket data for the validity of the future metrics.
Bug: 70571383
Test: statsd_test
Change-Id: Ib13c45574974e7b4e82bd8f305091dc93bda76f5
The key change is to revamp how we parse/store/match a log event, especially how we match repeated
field and attribution nodes, and how we construct dimensions and compare them.
+ We use a integer to encode the field of a log element. And also encode the FieldMatcher into an
integer and a bit mask. The log matching becomes 2 integer operations.
+ Dimension is stored as encoded field and value pair. Checking if 2 dimensions are equal is then
becoming checking if the underlying integers are equal. The integers are stored contiguously
in memory, so it's much faster than previous tree structure.
Start review from FieldValue.h
Test: statsd_test + new unit tests
Bug: 72659059
Change-Id: Iec8daeacdd3f39ab297c10ab9cd7b710a9c42e86
2/ Handle Subscription for alert.
3/ Support no_report_metric
Bug: 69522276
Test: all statsd unit tests passed.
Change-Id: I851b235f2d149b8602b0cad632d5bf541962f40a
+ Add log source whitelist in StatsdConfig
+ Some changes in UidMap API. Listener needs to be wp instead of sp.
+ Update dogfood app config to have log source
+ Increase the stats service thread pool size to 10 (9+1).
TODO: add unit tests(b/70805664). This unit test takes some time to write.
Test: statsd_test & manual
Change-Id: I129b1cc13db5114db7417580962bd7cc4438519d
We use one alarm clock for all pulled atoms.
If metrics from different configs are not aligned,
the clock will be set to repeat at higher and higher
frequency, and consume a lot of battery.
Current implementation assumes a 5min minimum bucket
size. New metric start time is set to be aligned to
the start time of statsd in the next 5min.
So it will ignore events up to 5min.
align puller alarm to minute bundary
Test: unit test
Change-Id: I77ffa3c13de363c780b1000181b9a9b780dd0846
The exception is EventMetricProducer. Each EventMetricProducer will still have a ProtoOutputStream
Because LogEvent comes as a fixed 4K, it's more memory efficient to have an 8k ProtoOutputStream for
storing the events.
Also removed finish() api in MetricProducer, which was intended to use with Dropbox.
Test: statsd_test & dogfood app
Bug: 70393808
Change-Id: I2efe4ecc76a88060a9aa5eb49d1fa6ea60bc5da8
Since there is a separate guardrail for memory used by uid map, we
no longer add the memory from uid map with the memory per each
config's metrics. We also prevent the byte size check from happening
too frequently. In order to mock the MetricsManager, we refactor
some of the existing methods.
Test: Added unit-tests and verified they all pass on marlin.
Change-Id: I15cf105f7d95f4016fdb0443b0a33eebe862cafb
