Docs

OTTER

OpenTracing Toolkit for ERlang

Build

OTTER uses rebar3 as build tool. It can be built with:

    rebar3 compile

However most likely you'll want to add it to your project in your build environment.

Dependencies

ibrowse HTTP client is used to send the HTTP/Thrift requests to Zipkin.

cowboy HTTP server (version 1.x API) is used to receive HTTP/Thrift (Zipkin binary format).

OpenTracing

OpenTracing is an open initiative to provide a set of terms and methods to produce, collect and correlate trace information in a distributed environment across different programming languages, platforms and protocols.

The concept defined for trace production is based on a span which is essentially a record of a handling in one environment. A span has a timestamp of when it started, a duration, a list of timestamped events marking the timing of important actions during the span and a list of key-value tags storing the parameters of the handled request (e.g. customer ids, transaction ids, results of subsequent actions). The span also contains id's to aid their correlation. The trace_id is used for correlating spans related to the handling of one request received from different systems. The trace_id is generated in the first system which starts handling a request (e.g. a frontend) and supposed to be passed on to other systems involved in the processing the same request. This is fairly simple when the protocols are fully under control and extensible (e.g. HTTP). Other id's recorded are the span_id and a parent_id referring to the parent span to help showing a hierarchical relationship of the spans in the trace collector.

After collecting this information, the span can be sent to a trace collector, which based on the id's of the received spans can correlate them and provide and end-to-end view of the request. Sending all produced spans could generate significant additional load on the system that produces them and also on the trace collector. It is recommended to filter the the spans before sending them to the collector.

The most mature trace collector at the time of the initial development is OpenZipkin. OTTER provides an interface to send spans to Zipkin using the HTTP/Thrift binary protocol.

The OpenTracing terminology defines information to be passed on across systems. The feasibility of this in most cases depends on the protocols used, and sometimes rather difficult to achieve. OTTER is not attempting to implement any of this functionality. It is possible though to initialize a span in OTTER with a trace_id and parent_id, but how these id's are passed across the systems is left to the particular implementation.

OTTER functionality

OTTER helps producing span information, filtering spans, sending to trace collector (Zipkin) and also counting and keeping a snapshot of the last occurrence of a span.

Producing span information

The main motivation behind the span collection of Otter is to make the instrumentation of existing code as simple as possible.

Types used in the API

The following type specifications are in otter.hrl

-type time_us() :: integer().           % timestamp in microseconds
-type info()    :: binary() | list() | atom() | integer().
-type ip4()     :: {integer(), integer(), integer(), integer()}.
-type service() :: binary() | list() | default | {binary() | list(), ip4(), integer()}.
-type trace_id():: integer().
-type span_id() :: integer().

-record(span , {
    timestamp   :: time_us(),           % timestamp of starting the span
    trace_id    :: trace_id(),          % 64 bit integer trace id
    name        :: info(),              % name of the span
    id          :: span_id(),           % 64 bit integer span id
    parent_id   :: span_id() | undefined, % 64 bit integer parent span id
    tags = []   :: [{info(), info()} | {info(), info(), service()}],  % span tags
    logs = []   :: [{time_us(), info()} | {info(), info(), service()}], % span logs
    duration    :: time_us()            % microseconds between span start/end
}).

-type span()    :: #span{}.

Functional API

Passing the span structure between requests, this API will probably require to pass the span in function calls if the function has something to add to the span. This requires more code changes, however when functions pass non-strict composite structures (e.g. maps or proplists) then inserting the span information is more or less trivial.

Start span with name only. Name should refer e.g. to the interface.

-spec span_start(info()) -> span().

Start span with name and trace_id where trace_id e.g. received from protocol.

-spec span_start(info(), integer()) -> span().

Start span with name, trace_id and parent span id e.g. received from protocol.

-spec span_start(info(), integer(), integer()) -> span().

Add a tag to the previously started span.

-spec span_tag(span(), info(), info()) -> span().

Add a tag to the previously started span with additional service information.

-spec span_tag(span(), info(), info(), service()) -> span().

Add a log/event to the previously started span

-spec span_log(span(), info()) -> span().

Add a log/event to the previously started span with additional service information0

-spec span_log(span(), info(), service()) -> span().

End span and invoke the span filter (see below)

-spec span_end(span()) -> ok.

Get span id's. Return the trace_id and the span id from the currently started span. This can be used e.g. when process "boundary" is to be passed and eventually new span needs this information. Also when these id's should be passed to a protocol interface for another system

-spec span_ids(span()) -> {trace_id(), span_id()}.

example :

    ...
    Span = otter:span_start("radius request"),
    ...
    ...
    Span1 = otter:span_tag(Span, "request_id", RequestId),
    ...
    ...
    Span2 = otter:span_plog(Span1, "invoke user db"),
    ...
    ...
    Span3 = otter:span_plog(Span2, "user db result"),
    Span4 = otter:span_ptag(Span3, "user db result", "ok"),
    ...
    ...
    Span5 = otter:span_ptag(Span4, "final result", "error"),
    Span6 = otter:span_ptag(Span5, "final result reason", "unknown user"),
    otter:span_pend(Span6),
    ...

Process API

The simplest API uses the process dictionary to store span information. This is probably the least work to implement in existing code.

Start span with name only. Name should refer e.g. to the interface.

-spec span_pstart(info()) -> ok.

Start span with name and trace_id where trace_id e.g. received from protocol.

-spec span_pstart(info(), trace_id()) -> ok.

Start span with name, trace_id and parent span id e.g. received from protocol.

-spec span_pstart(info(), trace_id(), span_id()) -> ok.

Add a tag to the previously started span.

-spec span_ptag(info(), info()) -> ok.

Add a tag to the previously started span with additional service information

-spec span_ptag(info(), info(), service()) -> ok.

Add a log/event to the previously started span

-spec span_plog(info()) -> ok.

Add a log/event to the previously started span with additional service information

-spec span_plog(info(), service()) -> ok.

End span and invoke the span filter (see below)

-spec span_pend() -> ok.

Get span id's. Return the trace_id and the span id from the currently started span. This can be used e.g. when process "boundary" is to be passed and eventually new span needs this information. Also when these id's should be passed to a protocol interface for another system

-spec span_pids() -> {trace_id(), span_id()}.

Return the current span. e.g. it can be handed to another process to continue collecting span information using the functional API.

-spec span_pget() -> span().

example :

    ...
    otter:span_pstart("radius request"),
    ...
    ...
    otter:span_ptag("request_id", RequestId),
    ...
    ...
    otter:span_plog("invoke user db"),
    ...
    ...
    otter:span_plog("user db result"),
    otter:span_ptag("user_db_result", "ok"),
    ...
    ...
    otter:span_ptag("final_result", "error"),
    otter:span_ptag("final_result_reason", "unknown user"),
    otter:span_pend(),
    ...

tag/log information

A note on the tag key/value and log types: the Zipkin interface requires string types. The Zipkin connector module (otter_conn_zipkin.erl) attempts to convert: integer, atom, list types to binary. This allows using these data types too as keys, values and logs, however any other "non-obvious" types would still need conversion (or simply can not be used). The impact of having an incompatible type in the tag keys, values and logs will likely crash the encoding/sending function in the Zipkin connector module resulting the loss of the particular batch of spans.

Adding service information to tags and logs means that otter adds a host structure to each of these elements. The extra optional service parameter in the relevant API calls can have 3 formats.

The atom default will include service/host information based on the following configuration parameters of the zipkin connector.

    ...
    {zipkin_tag_host_ip, {127,0,0,1}},
    {zipkin_tag_host_port, 0},
    {zipkin_tag_host_service, "otter_test"},
    ...

Name of the service as a string in binary() or list() format also adds the host information based on the configuration above, except the service name will be as specified in the parameter.

A 3 element tuple {Service, Ip, Port} which will be used to compose the information. This can be interesting to compose "ca" and "sa" opentracing tags where the host information may refer to a remote server/client node instead of the one where the span is generated.

Configuration

There is no configuration involved in the stage of producing span data. The paramers mentioned above are functionally specific to the zipkin connector. It was simpler to explain them though in this context.

Span Filtering

When the collection of span information is completed (i.e. span_pend or span_end/1 is called), filtering is invoked. Filtering is based on the tags collected in the span with the span name and the span duration added to the key/value pair list with keys : otter_span_name and otter_span_duration. The resulting key/value pair list which is used as input of the filter rules. With the examples above it can look like this :

    [
        {otter_span_name, "radius request"},
        {otter_span_duration, 1202},
        {"request_id", "6390266399200312"},
        {"user_db_result", "ok"},
        {"final_result", "error"},
        {"final_result_reason", "unknown user"}
    ]

This key/value pair list is passed to a sequence of conditions/actions pairs. In each pair the, conditions are a list of checks against the key/value pair list. If all conditions in the list are true, the actions are executed. An empty condition list always returns a positive match.

Filter conditions

Check the presence of a Key

    {present, Key}

Check whether 2 Keys have the same value

    {same, Key1, Key2}

Compare a value

The value of a Key/Value pair can be compared to a value

    {value, Key, ValueToCompare}

example: check the name of the span

    {value, otter_span_name, "radius request"}

Checking integer values

Key/Value pairs with integer values can be checked with the following conditions.

    {greater, Key, Integer}

    {less, Key, Integer}

    {between, Key, Integer1, Integer2}

example: check whether the span duration is greater than 5 seconds

    {greater, "otter_span_duration", 5000000}

Negate condition check

    {negate, Condition}

example: Check if the final result is other than ok

    {negate, {value, "final_result", "ok"}}

Filter Actions

Snapshot/Count

Snapshot/Count increases a counter with a key composed by a fixed prefix and values of Key/Values in the Key/Value list. The key is a list of parameters. Also it stores the last span that triggers the counter in an ets table. These cheap snapshots can be used for initial analysis of eventual problems. The snapshots and counter can be retrieved by the otter counter API (see below).

    {snapshot_count, Prefix, KeyList}

example: snapshot/count any request that take long in different counters for each span name and final result. The condition example above with the otter_span_duration could be used to trigger this action.

    {snapshot_count, [long_request], [otter_span_name, "final_result"]}

This will produce a counter and snapshot with e.g. such key :

    [long_request, "radius request", "ok"]

Send span to Zipkin

This action triggers sending the span to Zipkin

    send_to_zipkin

Stop evaluating further Condition/Action pairs

Normally each span triggers checking of all Condition/Action pairs in the sequence (executing all relevant actions). However if for a particular set of conditions it is not necessary, the break action can be used. When this is found in an Action list then all actions in the current list are executed and no further Condition/Action pairs are checked.

    break

Filter configuration

The filter rules are configured under filter_rules

Example

example Condition/Action (rule) list:

    [
        {
            %% Condition
            [
                {greater, "otter_span_duration", 5000000},
                {value, "otter_span_name", "radius request"}
            ],
            %% Action
            [
                {snap_count, [long_radius_request], []},
                send_to_zipkin
            ]
        },
        {
            %% Condition counts all requests with name and result
            [
            ],
            %% Action
            [
                {snap_count, [request], ["otter_span_name", "final_result"]}
            ]
        }

    ]

Sending a span to Zipkin

As a result of filter action send_to_zipkin the span is forwarded to the trace collector using HTTP/Thrift binary protocol. In the context of the span producing process the span is added to a buffer (ETS table). The content of this buffer is sent to Zipkin asynchronously in intervals configured in zipkin_batch_interval_ms (milliseconds).

The URI of the Zipkin trace collector is configured in zipkin_collector_uri.

Zipkin requires a node entry for every single tag collected in the span. This entry contains the service name and IP/Port of the node sending the span. If there is no tag/log is produced during span collecion with service/host paramer, the Zipkin connector module (otter_conn_zipkin.erl) can be configured to add an extra tag to each span during encoding the span by setting the zipkin_add_host_tag_to_span in the configuration. The value of the parameter should a tuple {Key, Value} which will be used as the default tag information. OpenZipkin uses the "lc" (Local Component) tag to display the service for a span.

example :

    {zipkin_add_host_tag_to_span, {"lc", ""}},

The default service/host information to be sent to zipkin is provided in zipkin_tag_host_service, zipkin_tag_host_ip and zipkin_tag_host_port configuration parameters.

It is also possible to add the default service/host information (as specified) in the configuration paramers above to each tag/log that does not have it explicitly set by setting the zipkin_add_default_service_to_logs and zipkin_add_default_service_to_tags to true. This however can increase the amount of data to be sent to zipkin significantly (depending on the number of logs/tags) so probably it is not recommended for most uses.

example :

    ...
    {zipkin_add_default_service_to_logs, false},
    {zipkin_add_default_service_to_tags, false},
    ...

Sending the span to Zipkin utilizes the ibrowse http client (which is the only dependency of OTTER).

Snapshot/Counter

As a result of the filter snap_count action, 2 ETS tables are used to count events (see snap_count action above) and in the same time store the last span information that has increased the counter. This can be considered a useful and fairly cheap troubleshooting tool.

Tho retrieve and manage these snapshot counters, OTTER provides API calls in the otter API module.

List counters

    -spec counter_list() -> [{list(), integer()}].

example:

3> otter:counter_list().
[{[long_span,test_request],1},
 {[otter_conn_zipkin,send_spans,failed],1},
 {[span_processed,"customer db lookup"],1},
 {[span_processed,test_request],1},
 {[long_span,"customer db lookup"],1}]

Retrieve snapshot for counter

    -spec counter_snapshot(list()) -> term().

example:

4> otter:counter_snapshot([long_span,test_request]).
[{[long_span,test_request],
  [{snap_timestamp,{2017,2,21,19,8,23,76525}},
   {data,{span,1487700503067982,3826404163842487863,
               test_request,1113017739039451686,undefined,
               [{customer_id,1},
                {transaction_id,3232323},
                {magic_tag,"wizz"},
                {magic_result,error},
                {db_result,"bad customer"},
                {final_result,error}],
               [{1487700503067998,"starting some magic"},
                {1487700503068000,"finished magic"},
                {1487700503068012,"db lookup"},
                {1487700503076491,"db lookup returned"}],
               8525}}]}]

Delete counter (and its snapshot)

    -spec counter_delete(list()) -> ok.

Delete all counters and snapshots

    -spec counter_delete_all() -> ok.

Built in HTTP/Thrift (zipkin protocol) server

OTTER has a HTTP/Thrift server that understands the Zipkin binary protocol. Originally was intended for testing purposes, but probably it can be generally useful to write own trace collector, or custom correlator for instrumented services.

The server is started up during otter application start if the server_zipkin_callback configuration parameter is set to a {Module, Function} tuple. The listening port can be specified with server_zipkin_port and defaults to 9411 if not defined. When spans are received on the server, they are decoded to otter #span{} record and individually handed to the specified callback. There is a stub/example callback module (otter_server_span_cb.erl) in the application. All the string types in the span are decoded to binary() types. i.e. it is not completely symmetrical with the encoder functions. While the encoder is "forgiving" when it comes to log text, tag key/values and other string parameters by accepting atom() | binary() | list() | integer() for convenience, obviously the decoder can not guess these.

The spans can be submitted to any URI path. The path is ignored by the server and the server always responds with HTTP response code 202 in the current implementation.

Example server configuration

    ...
    {server_zipkin_callback, {otter_server_span_cb, handle}},
    {server_zipkin_port, 19411},
    ...

OTTER Configuration

The OTTER application configuration is handled through the otter_config module (otter_config.erl). In the default implementation it uses the application environment. An example configuration can be found in the otter.app.src file.

    ...
    {http_client, ibrowse}, %% ibrowse | httpc
    {zipkin_collector_uri, "http://172.17.0.2:9411/api/v1/spans"},
    {zipkin_batch_interval_ms, 100},
    {zipkin_tag_host_ip, {127,0,0,1}},
    {zipkin_tag_host_port, 0},
    {zipkin_tag_host_service, "otter_test"},
    {zipkin_add_host_tag_to_span, {"lc", ""}},
    {zipkin_add_default_service_to_logs, false},
    {zipkin_add_default_service_to_tags, false},
    {server_zipkin_callback, {otter_server_span_cb, handle}},
    {server_zipkin_port, 19411},
    {filter_rules, [
        {
            [
                {greater, otter_span_duration, 1000}
            ],
            [
                {snapshot_count, [long_span], [otter_span_name]}
            ]
        },
        {
            [],
            [
                {snapshot_count, [span_processed], [otter_span_name]},
                send_to_zipkin
            ]
        }
    ]},
    ...

Acknowledgements

The development of ''otter'' was championed by Holger Winkelmann of Travelping. Both Travelping and bet365 kindly provided sponsorship for the initial development. The development was primarily done by Ferenc Holzhauser with design input from Chandru Mullaparthi.

License

Apache 2.0