Skip to main content
Version: 0.12

Metrics

One case you can deploy tremor for is to work with metrics. Not as a store but as a preprocessor, aggregator, and filter for metrics stores.

In this guide, we'll see how to use tremor for this. We'll set up an example where tremor batch, aggregates, and then forwards metrics. In addition, we'll take a look at tremor's metrics.

note

This example expects that you have some knowledge of tremor or went through the basics guide. We won't explain the concepts covered there again here.

In addition, we use docker compose for this guide as it requires additional software, and we want to avoid you having to install and configure a bunch of other services. We expect some familiarity with this as we will not explain the details.

All code for this guide can found on github.

Topics

This guide introduces the following new concepts.

  • codecs
  • Post- and Preprocessors
  • Connector configuraiton
  • pipeline operators
  • streams
  • aggregation
  • time-based windows
  • tremors internal metrics

Foundation

Let us start with the basic deployment:

  • influxdb for storage
  • chronograph for display
  • telegraf as an agent to collect some data
  • tremor for aggregation and filtering

A high level visualization of the application components:

graph LR A{telegraf UDP} -->|influx line protocol| B(tremor) B -->|Influx API| C{InfluxDB HTTP} D{Chronograf} --- C

The docker-compose file for this is here.

UDP Server

Something new here is that we fully define our connectors instead of using a pre-defined canned connector.

For the input, we use a udp_server connector. In the with clause, we can pass in the configuration.

Here are the fields of interest we have not discussed before.

codec is an option that can be passed to (nearly) all connectors, and it defines how data that arrives over this connection is decoded. In our case we use influx to decode the influxdb line protocol.

preprocessors are also an option that can be passed to (nearly) all connectors. You can use preprocessors to perform some processing on the byte stream that the connector provides. In our case, we use separate. This preprocessor will separate the byte stream with a given character or byte, by default, the \n newline.

note

Some connectors provide structured output instead of a byte stream. Structured connectors can't use a codec, preprocessors or postprocessor - after all, their output is already well defined. ::::

The config section holds a specific configuration to the connector type selected. Since we use the udp_server, we only have to specify the URL we want to listen on.

With this, we get:

define flow metrics
flow
# define the udp server
define connector upd_in from udp_server
with
# define the codec we use, in this case `influx` for the influx wire protocol
codec = "influx",
# define the preprocessors, we use separate to seperate events by lines
preprocessors = ["separate"],
# configure the connector itself, we listen to `0.0.0.0` on port `4242`
config = {
"url": "0.0.0.0:4242",
}
end;
end;

HTTP Client

On the other side we use a http_client. Since influxdb does not have a custom transport we are not forced to implement a influx connector. Instead we can use the influx codec and a http or udp connectors.

Connector naming follows a scheme here. Generally, something is named server if it accepts connections and client if it initiates connections. While there is a correlation between clients writing and servers reading, this is incidental and not a rule. A client connection could be reading and not a writing connector. :::

Our connector config here is slightly more elaborate. In addition to the URL, defining the target to write to, we also have a headers map that specifies the HTTP headers. This demonstrates nicely that more complex configurations are possible.

define flow metrics
flow
# define the udp server
define connector upd_in from udp_server
with
# define the codec we use, in this case `influx` for the influx wire protocol
codec = "influx",
# define the preprocessors, we use separate to seperate events by lines
preprocessors = ["separate"],
# configure the connector itself, we listen to `0.0.0.0` on port `4242`
config = {
"url": "0.0.0.0:4242",
}
end;

# define our http client
define connector influx_out from http_client
with
# we use the influx codec here as well
codec = "influx",
# configure the endpoint we're writing to
config = {
"url": "http://influxdb:8086/write?db=tremor",
# We use a custom header to identify that we're tremor
"headers": {"Client": ["Tremor"]}
}
end;
end;

deploy flow metrics;

Pipeline, wiring up and running

As in the basics tutorial, we will use the passthrough pipeline from the tremor::pipelines module. If you went through the basics guide, this is the same as before.

define flow metrics
flow
# import the pipeline module
use tremor::pipelines;
# define the udp server
define connector upd_in from udp_server
with
# define the codec we use, in this case `influx` for the influx wire protocol
codec = "influx",
# define the preprocessors, we use separate to seperate events by lines
preprocessors = ["separate"],
# configure the connector itself, we listen to `0.0.0.0` on port `4242`
config = {
"url": "0.0.0.0:4242",
}
end;

# define our http client
define connector influx_out from http_client
with
# we use the influx codec here as well
codec = "influx",
# configure the endpoint we're writing to
config = {
"url": "http://influxdb:8086/write?db=tremor",
# We use a custom header to identify that we're tremor
"headers": {"Client": ["Tremor"]}
}
end;

# Create the UDP server
create connector upd_in;

# Create the HTTP client
create connector influx_out;

# Create our pipeline
create pipeline passthrough from pipelines::passthrough;

# Connect the udp server to the pipeline
connect /connector/upd_in to /pipeline/passthrough;
# Connect the pipeline to the inflix client
connect /pipeline/passthrough to /connector/influx_out;

end;

# start our
deploy flow metrics;

Now with taht set you can grab the entire config from github and start it with docker-compose up.

You can find the chronograf UI at http://localhost:8888.

Batching

Tremor processes events one by one, meaning that each metric is considered its event. The upside of this is how events arrive. If it's one event per UDP message or 100 makes no difference. The downside is that InfluxDB requires events to be submitted in batches to maintain performance.

Since InfluxDB isn't the only destination that benefits from batched events, tremor provides a batching operator.

note

Batching using the batching operator is handled inside the connector. HTTP, for example, will combine a batch into a single request, but different connectors might handle this differently.

Adding the Batch operator

Operators are small pieces of logic that are written in rust. You can put them inside of a pipeline to perform specific actions. In our example, we use the batch operator.

Their use is similar to what you already know from pipelines and connectors. You first define operators with a set of configuration parameters, then create them for use.

We then use select to wire it up in the pipeline.

We use a maximum of 3000 events and a maximum delay of 5s for our operator.

Note that we select from in into batch and from batch into out, wiring up the event chain.

We also need to update our HTTP Client to use a postprocessor, namely the separate one, to join batch elements back together with newlines.

define flow metrics
flow

# define the udp server
define connector upd_in from udp_server
with
# define the codec we use, in this case `influx` for the influx wire protocol
codec = "influx",
# define the preprocessors, we use separate to seperate events by lines
preprocessors = ["separate"],
# configure the connector itself, we listen to `0.0.0.0` on port `4242`
config = {
"url": "0.0.0.0:4242",
}
end;


# define our http client
define connector influx_out from http_client
with
# we use the influx codec here as well
codec = "influx",
# define the postprocessor, we use separate to seperate events by lines
postprocessor = ["separate"],
# configure the endpoint we're writing to
config = {
"url": "http://influxdb:8086/write?db=tremor",
# We use a custom header to identify that we're tremor
"headers": {"Client": ["Tremor"]}
}
end;

# define our metrics pipeline to batch metrics
define pipeline metrics
pipeline
# define our batch operator.
# A batch will collect up to 3000 events, but never wait more then 5 seconds
# before emitting.
define operator batch from generic::batch
with
count = 3000,
timeout = nanos::from_seconds(5)
end;

create operator batch;

select event from in into batch;
select event from batch into out;
end;

# Create the udp server
create connector upd_in;

# Create the http client
create connector influx_out;

# Create our pipeline
create pipeline metrics from metrics;

# Connect the udp server to the pipeline
connect /connector/upd_in to /pipeline/metrics;
# Connect the pipeline to the inflix client
connect /pipeline/metrics to /connector/influx_out;

end;

# start our
deploy flow metrics;

Running

Now with taht set you can grab the entire config from github and start it with docker-compose up.

You can find the chronograf UI at http://localhost:8888.

Aggregation

Let's talk about aggregation. Sometimes it's helpful to aggregate some metrics before they're stored in a database. One reason could be to downsample the input. Another could be to create histograms for databases that don't natively support them.

note

The techniques used for aggregation aren't specific to metrics. You can apply the same methods to any event stream.

The aggregation we're looking for is "The histogram for each metric over a 10s and 1min window."

Grouping

To aggregate the metrics, we get in a useful we need to group our data by the "each metric" from our definition. Since we're using influx data as our input, a metric is identified by three fields:

  • The measurement
  • The tags
  • The field

Since influx combines multiple fields in a message, we first have to unroll this into each field being a single event.

::: note We use create stream here. Streams are just named nodes that have no own function or overhead. They serve as a way to chain up selects. :::

    # use types for checking if values are a number
use std::type;

# use record for exploding keys
use std::record;

# create our pre-aggregation stream
create stream aggregate;

# We change the structure of our event to be easier digestible for aggregation
select {
"measurement": event.measurement,
"tags": event.tags,
"field": group[2],
"value": event.fields[group[2]],
"timestamp": event.timestamp,
}
from in
group by set(event.measurement, event.tags, each(record::keys(event.fields)))
into aggregate
having type::is_number(event.value);

What we do here is for each key in fields, we create a metric, so if our record has two fields, we flatten that out and have two events. We then create a set of this field, the tags for the metric, and the measurement for the metric. This gives us a unique identifier for the metric/

This is needed to not confuse the values for different fields.

note

We use group[2] here, which refers to the second element of the current group (indexing starts at 0). For a string of the entire group, you can access n+1( in our example, since we have three elements, it would be group[3]).

Windowing

To aggregate over a time range, we use time-based tumbling windows. We can define them the same way we define everything else, a define statement. They are, however, not created. Instead, they're bound to the select statement they're used in.

    # use nanos for time
use std::time::nanos;

# define a 10 second window
define window `10secs` from tumbling
with
interval = nanos::from_seconds(10)
end;

# define a 1 minute window
define window `1min` from tumbling
with
interval = nanos::from_minutes(1)
end;

Aggregating

Once we have defined the windows, we can now use them to aggregate our data. We do this in a new select statement. After the from section, we have a square bracket and the window names. This syntax is what we call tilt frames. They're chained windows that will emit based on conditions.

We use the aggr::stats::hdr fimctopm that creates a histogram from the incoming data and outputs the given set of percentiles.

In addition we use the aggr::win::first to get the first timestamp.

note

We are using two windows here, a 10s one and a 1min one. The upside of using the tilt framing mechanism here is that we can do this without a loss in precision, as the 1min window is not an aggregate of six 10s windows but rather an aggregate of the raw data of them without having to duplicate it.

    # create a stream for normalization
create stream normalize;

select
{
"measurement": event.measurement,
"tags": patch event.tags of insert "window" => window end,
"stats": aggr::stats::hdr(event.value, [ "0.5", "0.9", "0.99", "0.999" ]),
"field": event.field,
"timestamp": aggr::win::first(event.timestamp), # we can't use min since it's a float
}
from aggregate[`10secs`, `1min`]
group by set(event.measurement, event.tags, event.field)
into normalize;

Normalisation

Last but not least we need to normalize this data basck to something the influx codec understands.

We can do this with another select statement.

select {
"measurement": event.measurement,
"tags": event.tags,
"fields": {
"count_#{event.field}": event.stats.count,
"min_#{event.field}": event.stats.min,
"max_#{event.field}": event.stats.max,
"mean_#{event.field}": event.stats.mean,
"stdev_#{event.field}": event.stats.stdev,
"var_#{event.field}": event.stats.var,
"p50_#{event.field}": event.stats.percentiles["0.5"],
"p90_#{event.field}": event.stats.percentiles["0.9"],
"p99_#{event.field}": event.stats.percentiles["0.99"],
"p99.9_#{event.field}": event.stats.percentiles["0.999"]
},
"timestamp": event.timestamp,
}
from normalize
into batch;

Running

Now with taht set you can grab the entire config from github and start it with docker-compose up.

You can find the chronograf UI at http://localhost:8888.

Tremor metrics

Finally, let us talk about how to access tremors own metrics. We choose to allow tremor to process them and forward them as required. This avoids an extra step for configuring metrics, implementing connectivity, or forcing the users into a specific way of monitoring.

This is done using the metrics connector.

Connectors

To enable metrics collection for connectors, we need to update their configuration. Namely we add the metrics_interval_s option. We choose 5 seconds to keep the overhead low.

  # define the udp server
define connector upd_in from udp_server
with
# define metrics interval
metrics_interval_s = 5,
# define the codec we use, in this case `influx` for the influx wire protocol
codec = "influx",
# define the preprocessors, we use separate to seperate events by lines
preprocessors = ["separate"],
# configure the connector itself, we listen to `0.0.0.0` on port `4242`
config = {
"url": "0.0.0.0:4242",
}
end;

# define our http client
define connector influx_out from http_client
with
# define metrics interval
metrics_interval_s = 5,
# we use the influx codec here as well
codec = "influx",
# define the postprocessors, we use separate to seperate events by lines
postprocessors = ["separate"],
# configure the endpoint we're writing to
config = {
"url": "http://influxdb:8086/write?db=tremor",
# We use a custom header to identify that we're tremor
"headers": {"Client": ["Tremor"]}
}
end;

Pipeline

The pipeline has to be set up as well. As pipelines don't come with a with section, it is configured as part of a configuration directive with the same name.

We choose the same 5-second interval.

  define connector metrics from metrics;

# define our metrics pipeline to batch metrics
define pipeline metrics
pipeline
#!config metrics_interval_s = 5
# ...

Metric Connector & Wiring

So with all our connectors and pipelines configured, we have to wire up the connector to the metrics pipeline. This works since, for internal metrics, we use the same structure as the influx codec.

define flow metrics
flow
# ...

# Define our hasher
define connector metrics from metrics;
# Create the internal metrics collector
create connector metrics;
# Connect the metrices to the pipeline
connect /connector/metrics to /pipeline/metrics;

#...

Now with taht set you can grab the entire config from github and start it with docker-compose up.

You can find the chronograf UI at http://localhost:8888.

Other backends

While this example is written using InfluxDB as a backend, it works equally with other backends.

TDengine

TDengine can quickly replace influx. The only difference is that we need to change the url in the http_client

A high-level visaulization of TDengine replacing InfluxDB via the Influx API:

graph LR A{telegraf UDP} -->|influx line protocol| B(tremor) B -->|Influx API| C{TDengine}

The connector definition for TDengine has a slightly different endpoint specification, but the application is otherwise the same:

  # define our http client
define connector influx_out from http_client
with
# define metrics interval
metrics_interval_s = 5,
# we use the influx codec here as well
codec = "influx",
# define the postprocessors, we use separate to seperate events by lines
postprocessors = ["separate"],
# configure the endpoint we're writing to
config = {
"url": "http://tdengine:6041/influxdb/v1/write?db=tremor&u=root&p=taosdata",
# We use a custom header to identify that we're tremor
"headers": {"Client": ["Tremor"]}
}
end;

Now with that set you can grab the entire config from github and start it with docker-compose up.

You can find the grafana UI at http://localhost:3000.

Note that you need to log in with user admin and password admin

QuestDB

QuestDB can also replace influx. There are a few differences however as Quest has constraints on column names and does not support the HTTP protocol for Influx Line Protocol. We can choose UDP based or TCP based distribution for QuestDB.

A high-level visaulization of QuestDB replacing InfluxDB via the Influx API:

graph LR A{telegraf UDP} -->|influx line protocol over UDP| B(tremor) B -->|Influx Line Protocol over TCP| C{QuestDB}

The connector definition for QuestDB has a slightly different endpoint specification and uses the TCP protocol for connectivity, but the configuration is otherwise the same:

  # define our http client
define connector influx_out from tcp_client
with
# define metrics interval
metrics_interval_s = 5,
# we use the influx codec here as well
codec = "influx",
# define the postprocessors, we use separate to seperate events by lines
postprocessors = ["separate"],
config = {
"url": "questdb:9009",
},
end;

We also need to replace some column names to conform to quest db constraints, we do this through the re::replace_all regular expression from tremor's standard library for convenience.

    # We change the structure of our event to be easier digestible for aggregation
select {
"measurement": re::replace_all("[ :\\./]", event.measurement, "_"), # QuestDB limitations
"tags": event.tags,
"field": group[2],
"value": event.fields[group[2]],
"timestamp": event.timestamp,
}
from in
group by set(event.measurement, event.tags, each(record::keys(event.fields)))
into aggregate
having type::is_number(event.value);

Now with that set you can grab the entire config from github and start it with docker-compose up.

You can find the grafana UI at http://localhost:3000. You can find the QuestDB Console UI at http://localhost:9000.

Note that you need to log in with user admin and password admin in the Grafana UI.