Integrate Exemplars with Python SDK

Author: cnnradams

docs/examples/basic_meter/view.py

@@ -82,7 +82,7 @@
 # parameter to specify the bucket ranges
 size_view = View(
     requests_size,
-    HistogramAggregator(config=[20, 40, 60, 80, 100]),
+    HistogramAggregator(config={"bounds": [20, 40, 60, 80, 100]}),
     label_keys=["environment"],  # is not used due to ViewConfig.UNGROUPED
     config=ViewConfig.UNGROUPED,
 )

docs/examples/exemplars/.ipynb_checkpoints/statistical_exemplars-checkpoint.ipynb

@@ -0,0 +1,40 @@
+OpenTelemetry Exemplars Example
+===============================
+
+Exemplars are example measurements for aggregations. While they are simple conceptually, exemplars can estimate any statistic about the input distribution, can provide links to sample traces for high latency requests, and much more.
+For more information about exemplars and how they work in OpenTelemetry, see the `spec <https://github.com/open-telemetry/oteps/pull/113>`_
+
+Examples
+--------
+
+Installation
+
+.. code-block:: sh
+
+    pip install opentelemetry-api
+    pip install opentelemetry-sdk
+    pip install matplotlib # may have to install Qt as well
+    pip install numpy
+
+    pip install opentelemetry-exporter-cloud-monitoring # if you want to export exemplars to cloud monitoring
+
+Statistical exemplars
+^^^^^^^^^^^^^^^^^^^^^
+
+The opentelemetry SDK provides a way to sample exemplars statistically:
+
+    - Exemplars will be picked to represent the input distribution, without unquantifiable bias
+    - A "sample_count" attribute will be set on each exemplar to quantify how many measurements each exemplar represents
+
+See 'statistical_exemplars.ipynb' for the example (TODO: how do I link this?)
+
+Semantic exemplars
+^^^^^^^^^^^^^^^^^^
+
+Semantic exemplars are exemplars that have not been sampled statistically,
+but instead aim to provide value as individual exemplars.
+They will have a trace id/span id attached for the active trace when the exemplar was recorded,
+and they may focus on measurements with abnormally high/low values.
+
+'semantic_exemplars.py' shows how to generate exemplars for a histogram aggregation.
+Currently only the Google Cloud Monitoring exporter supports uploading these exemplars.

docs/examples/exemplars/README.rst

@@ -0,0 +1,67 @@
+# Copyright The OpenTelemetry Authors
+#
+# 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.
+#
+"""
+This example shows how to generate "semantic" exemplars for a histogram, and how to export them to Google Cloud Monitoring.
+"""
+
+import random
+import time
+
+from opentelemetry import metrics
+from opentelemetry.sdk.metrics import (
+    MeterProvider,
+    ValueRecorder,
+)
+from opentelemetry.sdk.metrics.export import ConsoleMetricsExporter
+from opentelemetry.sdk.metrics.export.aggregate import (
+    HistogramAggregator,
+)
+from opentelemetry.sdk.metrics.view import View, ViewConfig
+
+# Set up OpenTelemetry metrics
+metrics.set_meter_provider(MeterProvider(stateful=False))
+meter = metrics.get_meter(__name__)
+
+# Use the Google Cloud Monitoring Metrics Exporter since its the only one that currently supports exemplars
+metrics.get_meter_provider().start_pipeline(meter, ConsoleMetricsExporter(), 10)
+
+# Create our duration metric
+request_duration = meter.create_metric(
+    name="request_duration",
+    description="duration (ms) of incoming requests",
+    unit="ms",
+    value_type=int,
+    metric_type=ValueRecorder,
+)
+
+# Add a Histogram view to our duration metric, and make it generate 1 exemplars per bucket
+duration_view = View(
+    request_duration,
+    # Latency in buckets:
+    # [>=0ms, >=25ms, >=50ms, >=75ms, >=100ms, >=200ms, >=400ms, >=600ms, >=800ms, >=1s, >=2s, >=4s, >=6s]
+    # We want to generate 1 exemplar per bucket, where each exemplar has a linked trace that was recorded.
+    # So we need to set num_exemplars to 1 and not specify statistical_exemplars (defaults to false)
+    HistogramAggregator(config={"bounds": [0, 25, 50, 75, 100, 200, 400, 600, 800, 1000, 2000, 4000, 6000],
+        "num_exemplars": 1}),
+    label_keys=["environment"],
+    config=ViewConfig.LABEL_KEYS,
+)
+
+meter.register_view(duration_view)
+
+for i in range(100):
+    # Generate some random data for the histogram with a dropped label "customer_id"
+    request_duration.record(random.randint(1, 8000), {"environment": "staging", "customer_id": random.randint(1, 100)})
+    time.sleep(1)

docs/examples/exemplars/semantic_exemplars.py

@@ -0,0 +1,131 @@
+import numpy as np
+import matplotlib.pyplot as plt
+import random
+
+from collections import defaultdict
+
+from opentelemetry import metrics
+from opentelemetry.sdk.metrics import Counter, MeterProvider
+from opentelemetry.sdk.metrics.export.aggregate import SumAggregator
+from opentelemetry.sdk.metrics.export.controller import PushController
+from opentelemetry.sdk.metrics.export.in_memory_metrics_exporter import InMemoryMetricsExporter
+from opentelemetry.sdk.metrics.view import View, ViewConfig
+
+## set up opentelemetry
+
+# Sets the global MeterProvider instance
+metrics.set_meter_provider(MeterProvider())
+
+meter = metrics.get_meter(__name__)
+
+# Export to a python list so we can do stats with the data
+exporter = InMemoryMetricsExporter()
+
+# instead of waiting for the controller to tick over time, we will just tick it ourselves
+controller = PushController(meter, exporter, 500)
+
+# Create the metric that we will use
+bytes_counter = meter.create_metric(
+    name="bytes_counter",
+    description="Number of bytes received by service",
+    unit="By",
+    value_type=int,
+    metric_type=Counter,
+)
+
+# Every time interval we will collect 100 exemplars statistically (selected without bias)
+aggregator_config = {"num_exemplars": 100, "statistical_exemplars": True}
+
+# Assign a Sum aggregator to `bytes_counter` that collects exemplars
+counter_view = View(
+    bytes_counter,
+    SumAggregator(config=aggregator_config),
+    label_keys=["environment"],
+    config=ViewConfig.LABEL_KEYS,
+)
+
+meter.register_view(counter_view)
+
+## generate the random metric data
+
+def unknown_customer_calls():
+    """Generate customer call data to our application"""
+
+    # set a random seed for consistency of data for example purposes
+    np.random.seed(1)
+    # Make exemplar selection consistent for example purposes
+    random.seed(1)
+
+    # customer 123 is a big user, and made 1000 requests in this timeframe
+    requests = np.random.normal(1000, 250, 1000) # 1000 requests with average 1000 bytes, covariance 100
+
+    for request in requests:
+        bytes_counter.add(int(request), {"environment": "production", "method": "REST", "customer_id": 123})
+
+    # customer 247 is another big user, making fewer, but bigger requests
+    requests = np.random.normal(5000, 1250, 200) # 200 requests with average size of 5k bytes
+
+    for request in requests:
+        bytes_counter.add(int(request), {"environment": "production", "method": "REST", "customer_id": 247})
+
+    # There are many other smaller customers
+    for customer_id in range(250):
+        requests = np.random.normal(1000, 250, np.random.randint(1, 10))
+        method = "REST" if np.random.randint(2) else "gRPC"
+        for request in requests:
+            bytes_counter.add(int(request), {"environment": "production", "method": method, "customer_id": customer_id})
+
+unknown_customer_calls()
+
+# Tick the controller so it sends metrics to the exporter
+controller.tick()
+
+# collect metrics from our exporter
+metric_data = exporter.get_exported_metrics()
+
+# get the exemplars from the bytes_in counter aggregator
+aggregator = metric_data[0].aggregator
+exemplars = aggregator.checkpoint_exemplars
+
+# Sum up the total bytes in per customer from all of the exemplars collected
+customer_bytes_map = defaultdict(int)
+for exemplar in exemplars:
+    customer_bytes_map[exemplar.dropped_labels] += exemplar.value
+
+
+customer_bytes_list = sorted(list(customer_bytes_map.items()), key=lambda t: t[1], reverse=True)
+
+# Save our top 5 customers and sum all of the rest into "Others".
+top_5_customers = [("Customer {}".format(dict(val[0])["customer_id"]), val[1]) for val in customer_bytes_list[:5]] + [("Other Customers", sum([val[1] for val in customer_bytes_list[5:]]))]
+
+# unzip the data into X (sizes of each customer's contribution) and labels
+labels, X = zip(*top_5_customers)
+
+# create the chart with matplotlib and show it
+plt.pie(X, labels=labels)
+plt.show()
+
+# Estimate how many bytes customer 123 sent
+customer_123_bytes = customer_bytes_map[(("customer_id", 123), ("method", "REST"))]
+
+# Since the exemplars were randomly sampled, all sample_counts will be the same
+sample_count = exemplars[0].sample_count
+print("sample count", sample_count, "custmer", customer_123_bytes)
+full_customer_123_bytes = sample_count * customer_123_bytes
+
+# With seed == 1 we get 1008612 - quite close to the statistical mean of 1000000! (more exemplars would make this estimation even more accurate)
+print("Customer 123 sent about {} bytes this interval".format(int(full_customer_123_bytes)))
+
+# Determine the top 25 customers by how many bytes they sent in exemplars
+top_25_customers = customer_bytes_list[:25]
+
+# out of those 25 customers, determine how many used grpc, and come up with a ratio
+percent_grpc = len(list(filter(lambda customer_value: customer_value[0][1][1] == "gRPC", top_25_customers))) / len(top_25_customers)
+
+print("~{}% of the top 25 customers (by bytes in) used gRPC this interval".format(int(percent_grpc*100)))
+
+# Determine the 50th, 90th, and 99th percentile of byte size sent in
+quantiles = np.quantile([exemplar.value for exemplar in exemplars], [0.5, 0.9, 0.99])
+print("50th Percentile Bytes In:", int(quantiles[0]))
+print("90th Percentile Bytes In:", int(quantiles[1]))
+print("99th Percentile Bytes In:", int(quantiles[2]))

docs/examples/exemplars/statistical_exemplars.ipynb

@@ -77,11 +77,12 @@ def export(
     ) -> "MetricsExportResult":
         for record in metric_records:
             print(
-                '{}(data="{}", labels="{}", value={})'.format(
+                '{}(data="{}", labels="{}", value={}, exemplars={})'.format(
                     type(self).__name__,
                     record.instrument,
                     record.labels,
                     record.aggregator.checkpoint,
+                    record.aggregator.checkpoint_exemplars
                 )
             )
         return MetricsExportResult.SUCCESS