event-metrics 0.2.0

An embedded, event-time metric collection library

Event Metrics

An embedded, event-time metric collection library built for serving system

Metric systems like prometheus aggregate metric at “processing time”, whenever the scraper is able to scrape the metric. event_metrics capture and record metrics at event time, when the event happens.

Features

Comparing to other metric system, the event_metrics library:

• Write data to sqlite3 database to keep low-memory footprint

• Minimal dependency (only requires numpy for numeric aggregation)

• Aggregate with the full data by default (no reservior sampling)

• Allow select from past duration with timedelta windowing

• Timestamp all observation by default

• Small API footprint: observe and query is all you need to know

• Compute raw timeseries with different aggregation strategies:

  • Scalers: last, min, max, mean, count, sum

  • Buckets for histogram

  • Percentiles for summary

  • Array and timestamps for native python wrangling

• Metrics can be labeled with arbitrary key value pair and querying supports multidimensional label matching.

Install

• Install from source: pip install -e .

• PyPI package is work in progress

Usage

from event_metrics import MetricConnection

conn = MetricConnection("/tmp/event_metrics_demo")

conn.observe("latency", 1.2)
conn.increment("counter", -1)

# labeling
conn.observe("latency", 2.0, labels={"service": "myapp", "other": "label"})

# querying
(conn.query("latency", labels={"service":"myapp"})
      # select from past duration using one of the following
     .from_beginning()
     .from_timestamp(...)
     .from_timedelta(...)

      # perform aggregation using one of the following
     .to_scaler(agg="last/min/max/mean/count/sum") # counter, guage
     .to_buckets(buckets=[1, 5, 10], cumulative=False) # histogram
     .to_percentiles(percnetiles=[50, 90, 99]) # summary
     .to_array() # -> value array
     .to_timestamps() # -> timestamp array
     .to_timestamps_array() # -> 2 array, (timestamp, value array)
)

Speed

The library is fast enough. It can ingest about 34,000 data point per seconds:

You can run pytest tests -k bench to generate benchmark on local hardware:

-------------------------------------------------- benchmark: 1 tests --------------------------------------------------
Name (time in us)            Min       Max     Mean  StdDev   Median     IQR  Outliers  OPS (Kops/s)  Rounds  Iterations
------------------------------------------------------------------------------------------------------------------------
test_bench_ingestion     25.3340  297.9320  28.8541  9.1582  26.8090  0.8650   521;814       34.6571    6496           1
------------------------------------------------------------------------------------------------------------------------