Carbon-like graphite line mode relay.

This project aims to be a fast replacement of the original Carbon relay

The main reason to build a replacement is performance and configurability. Carbon is single threaded, and sending metrics to multiple consistent-hash clusters requires chaining of relays. This project provides a multithreaded relay which can address multiple targets and clusters for each and every metric based on pattern matches.

There are a couple more replacement projects out there we know of, which are carbon-relay-ng and graphite-relay

Compared to carbon-relay-ng, this project does provide carbon's consistent-hash routing. graphite-relay, which does this, however doesn't do metric-based matches to direct the traffic, which this project does as well. To date, carbon-c-relay can do aggregations, failover targets and more.

The relay is a simple program that reads its routing information from a file. The command line arguments allow to set the location for this file, as well as the amount of dispatchers (worker threads) to use for reading the data from incoming connections and passing them onto the right destination(s). The route file supports two main constructs: clusters and matches. The first define groups of hosts data metrics can be sent to, the latter define which metrics should be sent to which cluster. Aggregation rules are seen as matches.

For every metric received by the relay, cleansing is performed. The following changes are performed before any match, aggregate or rewrite rule sees the metric:

• double dot elimination (necessary for correctly functioning consistent hash routing)
• trailing/leading dot elimination
• whitespace normalisation (this mostly affects output of the relay to other targets: metric, value and timestamp will be separated by a single space only, ever)
• irregular char replacement with underscores (_), currently irregular is defined as not being in [0-9a-zA-Z-_:#], but can be overridden on the command line.

The route file syntax is as follows:

# comments are allowed in any place and start with a hash (#)

cluster <name>
    <forward | any_of [useall] | failover | <carbon_ch | fnv1a_ch | jump_fnv1a_ch> [replication <count>]>
        <host[:port][=instance] [proto <udp | tcp>]> ...
    ;
cluster <name>
    file [ip]
        </path/to/file> ...
    ;
match
        <* | expression ...>
    send to <cluster ... | blackhole>
    [stop]
    ;
rewrite <expression>
    into <replacement>
    ;
aggregate
        <expression> ...
    every <interval> seconds
    expire after <expiration> seconds
    [timestamp at <start | middle | end> of bucket]
    compute <sum | count | max | min | average |
             median | percentile<%> | variance | stddev> write to
        <metric>
    [compute ...]
    [send to <cluster ...>]
    [stop]
    ;

Multiple clusters can be defined, and need not to be referenced by a match rule. All clusters point to one or more hosts, except the file cluster which writes to files in the local filesystem. host may be an IPv4 or IPv6 address, or a hostname. Since host is followed by an optional : and port, for IPv6 addresses not to be interpreted wrongly, either a port must be given, or the IPv6 address surrounded by brackets, e.g. [::1]. An optional proto udp or proto tcp may be added to specify the use of UDP or TCP to connect to the remote server. When omitted this defaults to a TCP connection.

The forward and file clusters simply send everything they receive to all defined members (host addresses or files). The any_of cluster is a small variant of the forward cluster, but instead of sending to all defined members, it sends each incoming metric to one of defined members. This is not much useful in itself, but since any of the members can receive each metric, this means that when one of the members is unreachable, the other members will receive all of the metrics. This can be useful when the cluster points to other relays. The any_of router tries to send the same metrics consistently to the same destination. The failover cluster is like the any_of cluster, but sticks to the order in which servers are defined. This is to implement a pure failover scenario between servers. The carbon_ch cluster sends the metrics to the member that is responsible according to the consistent hash algorithm (as used in the original carbon), or multiple members if replication is set to more than 1. The fnv1a_ch cluster is a identical in behaviour to carbon_ch, but it uses a different hash technique (FNV1a) which is faster but more importantly defined to get by a limitation of carbon_ch to use both host and port from the members. This is useful when multiple targets live on the same host just separated by port. The instance that original carbon uses to get around this can be set by appending it after the port, separated by an equals sign, e.g. 127.0.0.1:2006=a for instance a. When using the fnv1a_ch cluster, this instance overrides the hash key in use. This allows for many things, including masquerading old IP addresses, but mostly to make the hash key location to become agnostic of the (physical) location of that key. For example, usage like 10.0.0.1:2003=4d79d13554fa1301476c1f9fe968b0ac would allow to change port and/or ip address of the server that receives data for the instance key. Obviously, this way migration of data can be dealt with much more conveniently. The jump_fnv1a_ch cluster is also a consistent hash cluster like the previous two, but it does not take the server information into account at all. Whether this is useful to you depends on your scenario. The jump hash has a much better balancing over the servers defined in the cluster, at the expense of not being able to remove any server but the last in order. What this means is that this hash is fine to use with ever growing clusters where older nodes are also replaced at some point. If you have a cluster where removal of old nodes takes place often, the jump hash is not suitable for you. Jump hash works with servers in an ordered list without gaps. To influence the ordering, the instance given to the server will be used as sorting key. Without, the order will be as given in the file. It is a good practice to fix the order of the servers with instances such that it is explicit what the right nodes for the jump hash are.

DNS hostnames are resolved to a single address, according to the preference rules in RFC 3484. The any_of cluster has an explicit useall flag that enables a hostname to resolve to multiple addresses. Each address returned becomes a cluster destination.

Match rules are the way to direct incoming metrics to one or more clusters. Match rules are processed top to bottom as they are defined in the file. It is possible to define multiple matches in the same rule. Each match rule can send data to one or more clusters. Since match rules "fall through" unless the stop keyword is added, carefully crafted match expression can be used to target multiple clusters or aggregations. This ability allows to replicate metrics, as well as send certain metrics to alternative clusters with careful ordering and usage of the stop keyword. The special cluster blackhole discards any metrics sent to it. This can be useful for weeding out unwanted metrics in certain cases. Because throwing metrics away is pointless if other matches would accept the same data, a match with as destination the blackhole cluster, has an implicit stop.

Rewrite rules take a regular input to match incoming metrics, and transform them into the desired new metric name. In the replacement, backreferences are allowed to match capture groups defined in the input regular expression. A match of server\.(x|y|z)\. allows to use e.g. role.\1. in the substitution. A few caveats apply to the current implementation of rewrite rules. First, their location in the config file determines when the rewrite is performed. The rewrite is done in-place, as such a match rule before the rewrite would match the original name, a match rule after the rewrite no longer matches the original name. Care should be taken with the ordering, as multiple rewrite rules in succession can take place, e.g. a gets replaced by b and b gets replaced by c in a succeeding rewrite rule. The second caveat with the current implementation, is that the rewritten metric names are not cleansed, like newly incoming metrics are. Thus, double dots and potential dangerous characters can appear if the replacement string is crafted to produce them. It is the responsibility of the writer to make sure the metrics are clean. If this is an issue for routing, one can consider to have a rewrite-only instance that forwards all metrics to another instance that will do the routing. Obviously the second instance will cleanse the metrics as they come in. The backreference notation allows to lowercase and uppercase the replacement string with the use of the underscore (_) and carret (^) symbols following directly after the backslash. For example, role.\_1. as substitution will lowercase the contents of \1.

The aggregations defined take one or more input metrics expressed by one or more regular expresions, similar to the match rules. Incoming metrics are aggregated over a period of time defined by the interval in seconds. Since events may arrive a bit later in time, the expiration time in seconds defines when the aggregations should be considered final, as no new entries are allowed to be added any more. On top of an aggregation multiple aggregations can be computed. They can be of the same or different aggregation types, but should write to a unique new metric. The metric names can include back references like in rewrite expressions, allowing for powerful single aggregation rules that yield in many aggregations. When no send to clause is given, produced metrics are sent to the relay as if they were submitted from the outside, hence match and aggregation rules apply to those. Care should be taken that loops are avoided this way. For this reason, the use of the send to clause is encouraged, to direct the output traffic where possible. Like for match rules, it is possible to define multiple cluster targets. Also, like match rules, the stop keyword applies to control the flow of metrics in the matching process.

Carbon-c-relay evolved over time, growing features on demand as the tool proved to be stable and fitting the job well. Below follow some annotated examples of constructs that can be used with the relay.

Clusters can be defined as much as necessary. They receive data from match rules, and their type defines which members of the cluster finally get the metric data. The simplest cluster form is a forward cluster:

cluster send-through
    forward
        10.1.0.1
    ;

Any metric sent to the send-through cluster would simply be forwarded to the server at IPv4 address 10.1.0.1. If we define multiple servers, all of those servers would get the same metric, thus:

cluster send-through
    forward
        10.1.0.1
        10.2.0.1
    ;

The above results in a duplication of metrics send to both machines. This can be useful, but most of the time it is not. The any_of cluster type is like forward, but it sends each incoming metric to any of the members. The same example with such cluster would be:

cluster send-to-any-one
    any_of 10.1.0.1:2010 10.1.0.1:2011;

This would implement a multipath scenario, where two servers are used, the load between them is spread, but should any of them fail, all metrics are sent to the remaining one. This typically works well for upstream relays, or for balancing carbon-cache processes running on the same machine. Should any member become unavailable, for instance due to a rolling restart, the other members receive the traffic. If it is necessary to have true fail-over, where the secondary server is only used if the first is down, the following would implement that:

cluster try-first-then-second
    failover 10.1.0.1:2010 10.1.0.1:2011;

These types are different from the two consistent hash cluster types:

cluster graphite
    carbon_ch
        127.0.0.1:2006=a
        127.0.0.1:2007=b
        127.0.0.1:2008=c
    ;

If a member in this example fails, all metrics that would go to that member are kept in the queue, waiting for the member to return. This is useful for clusters of carbon-cache machines where it is desirable that the same metric ends up on the same server always. The carbon_ch cluster type is compatible with carbon-relay consistent hash, and can be used for existing clusters populated by carbon-relay. For new clusters, however, it is better to use the fnv1a_ch cluster type, for it is faster, and allows to balance over the same address but different ports without an instance number, in constrast to carbon_ch.

Because we can use multiple clusters, we can also replicate without the use of the forward cluster type, in a more intelligent way:

cluster dc-old
    carbon_ch replication 2
        10.1.0.1
        10.1.0.2
        10.1.0.3
    ;
cluster dc-new1
    fnv1a_ch replication 2
        10.2.0.1
        10.2.0.2
        10.2.0.3
    ;
cluster dc-new2
    fnv1a_ch replication 2
        10.3.0.1
        10.3.0.2
        10.3.0.3
    ;

match *
    send to dc-old
    ;
match *
    send to
        dc-new1
        dc-new2
    stop
    ;

In this example all incoming metrics are first sent to dc-old, then dc-new1 and finally to dc-new2. Note that the cluster type of dc-old is different. Each incoming metric will be send to 2 members of all three clusters, thus replicating to in total 6 destinations. For each cluster the destination members are computed independently. Failure of clusters or members does not affect the others, since all have individual queues. The above example could also be written using three match rules for each dc, or one match rule for all three dcs. The difference is mainly in performance, the number of times the incoming metric has to be matched against an expression. The stop rule in dc-new match rule is not strictly necessary in this example, because there are no more following match rules. However, if the match would target a specific subset, e.g. ^sys\., and more clusters would be defined, this could be necessary, as for instance in the following abbreviated example:

cluster dc1-sys ... ;
cluster dc2-sys ... ;

cluster dc1-misc ... ;
cluster dc2-misc ... ;

match ^sys\. send to dc1-sys;
match ^sys\. send to dc2-sys stop;

match * send to dc1-misc;
match * send to dc2-misc stop;

As can be seen, without the stop in dc2-sys' match rule, all metrics starting with sys. would also be send to dc1-misc and dc2-misc. It can be that this is desired, of course, but in this example there is a dedicated cluster for the sys metrics.

Suppose there would be some unwanted metric that unfortunately is generated, let's assume some bad/old software. We don't want to store this metric. The blackhole cluster is suitable for that, when it is harder to actually whitelist all wanted metrics. Consider the following:

match
        some_legacy1$
        some_legacy2$
    send to blackhole
    stop;

This would throw away all metrics that end with some_legacy, that would otherwise be hard to filter out. Since the order matters, it can be used in a construct like this:

cluster old ... ;
cluster new ... ;

match * send to old;

match unwanted send to blackhole stop;

match * send to new;

In this example the old cluster would receive the metric that's unwanted for the new cluster. So, the order in which the rules occur does matter for the execution.

The relay is capable of rewriting incoming metrics on the fly. This process is done based on regular expressions with capture groups that allow to substitute parts in a replacement string. Rewrite rules allow to cleanup metrics from applications, or provide a migration path. In it's simplest form a rewrite rule looks like this:

rewrite ^server\.(.+)\.(.+)\.([a-zA-Z]+)([0-9]+)
    into server.\_1.\2.\3.\3\4
    ;

In this example a metric like server.DC.role.name123 would be transformed into server.dc.role.name.name123. For rewrite rules hold the same as for matches, that their order matters. Hence to build on top of the old/new cluster example done earlier, the following would store the original metric name in the old cluster, and the new metric name in the new cluster:

match * send to old;

rewrite ... ;

match * send to new;

Note that after the rewrite, the original metric name is no longer available, as the rewrite happens in-place.

Aggregations are probably the most complex part of carbon-c-relay. Two ways of specifying aggregates are supported by carbon-c-relay. The first, static rules, are handled by an optimiser which tries to fold thousands of rules into groups to make the matching more efficient. The second, dynamic rules, are very powerful compact definitions with possibly thousands of internal instantiations. A typical static aggregation looks like:

aggregate
        ^sys\.dc1\.somehost-[0-9]+\.somecluster\.mysql\.replication_delay
        ^sys\.dc2\.somehost-[0-9]+\.somecluster\.mysql\.replication_delay
    every 10 seconds
    expire after 35 seconds
    timestamp at end of bucket
    compute sum write to
        mysql.somecluster.total_replication_delay
    compute average write to
        mysql.somecluster.average_replication_delay
    compute max write to
        mysql.somecluster.max_replication_delay
    compute count write to
        mysql.somecluster.replication_delay_metric_count
    ;

In this example, four aggregations are produced from the incoming matching metrics. In this example we could have written the two matches as one, but for demonstration purposes we did not. Obviously they can refer to different metrics, if that makes sense. The every 10 seconds clause specifies in what interval the aggregator can expect new metrics to arrive. This interval is used to produce the aggregations, thus each 10 seconds 4 new metrics are generated from the data received sofar. Because data may be in transit for some reason, or generation stalled, the expire after clause specifies how long the data should be kept before considering a data bucket (which is aggregated) to be complete. In the example, 35 was used, which means after 35 seconds the first aggregates are produced. It also means that metrics can arrive 35 seconds late, and still be taken into account. The exact time at which the aggregate metrics are produced is random between 0 and interval (10 in this case) seconds after the expiry time. This is done to prevent thundering herds of metrics for large aggregation sets. The timestamp that is used for the aggregations can be specified to be the start, middle or end of the bucket. Original carbon-aggregator.py uses start, while carbon-c-relay's default has always been end. The compute clauses demonstrate a single aggregation rule can produce multiple aggregates, as often is the case. Internally, this comes for free, since all possible aggregates are always calculated, whether or not they are used. The produced new metrics are resubmitted to the relay, hence matches defined before in the configuration can match output of the aggregator. It is important to avoid loops, that can be generated this way. In general, splitting aggregations to their own carbon-c-relay instance, such that it is easy to forward the produced metrics to another relay instance is a good practice.

The previous example could also be written as follows to be dynamic:

aggregate
        ^sys\.dc[0-9].(somehost-[0-9]+)\.([^.]+)\.mysql\.replication_delay
    every 10 seconds
    expire after 35 seconds
    compute sum write to
        mysql.host.\1.replication_delay
    compute sum write to
        mysql.host.all.replication_delay
    compute sum write to
        mysql.cluster.\2.replication_delay
    compute sum write to
        mysql.cluster.all.replication_delay
    ;

Here a single match, results in four aggregations, each of a different scope. In this example aggregation based on hostname and cluster are being made, as well as the more general all targets, which in this example have both identical values. Note that with this single aggregation rule, both per-cluster, per-host and total aggregations are produced. Obviously, the input metrics define which hosts and clusters are produced.

With use of the send to clause, aggregations can be made more intuitive and less error-prone. Consider the below example:

cluster graphite fnv1a_ch ip1 ip2 ip3;

aggregate ^sys\.somemetric
    every 60 seconds
    expire after 75 seconds
    compute sum write to
        sys.somemetric
    send to graphite
    stop
    ;

match * send to graphite;

It sends all incoming metrics to the graphite cluster, except the sys.somemetric ones, which it replaces with a sum of all the incoming ones. Without a stop in the aggregate, this causes a loop, and without the send to, the metric name can't be kept its original name, for the output now directly goes to the cluster.

When carbon-c-relay is run without -d or -s arguments, statistics will be produced and sent to the relay itself in the form of carbon.relays.<hostname>.*. The hostname is determined on startup, and can be overriden using the -H argument. While many metrics have a similar name to what carbon-cache.py would produce, their values are different. To obtain a more compatible set of values, the -m argument can be used to make values non-cumulative, that is, they will report the change compared to the previous value. By default, most values are running counters which only increase over time. The use of the nonNegativeDerivative() function from graphite is useful with these. The default sending interval is 1 minute (60 seconds), but can be overridden using the -S argument specified in seconds.

The following metrics are produced in the carbon.relays.<hostname> namespace:

• metricsReceived

  The number of metrics that were received by the relay. Received here means that they were seen and processed by any of the dispatchers.

• metricsSent

  The number of metrics that were sent from the relay. This is a total count for all servers combined. When incoming metrics are duplicated by the cluster configuration, this counter will include all those duplications. In other words, the amount of metrics that were successfully sent to other systems. Note that metrics that are processed (received) but still in the sending queue (queued) are not included in this counter.

• metricsQueued

  The total number of metrics that are currently in the queues for all the server targets. This metric is not cumulative, for it is a sample of the queue size, which can (and should) go up and down. Therefore you should not use the derivative function for this metric.

• metricsDropped

  The total number of metric that had to be dropped due to server queues overflowing. A queue typically overflows when the server it tries to send its metrics to is not reachable, or too slow in ingesting the amount of metrics queued. This can be network or resource related, and also greatly depends on the rate of metrics being sent to the particular server.

• metricsBlackholed

  The number of metrics that did not match any rule, or matched a rule with blackhole as target. Depending on your configuration, a high value might be an indication of a misconfiguration somewhere. These metrics were received by the relay, but never sent anywhere, thus they disappeared.

• metricStalls

  The number of times the relay had to stall a client to indicate that the downstream server cannot handle the stream of metrics. A stall is only performed when the queue is full and the server is actually receptive of metrics, but just too slow at the moment. Stalls typically happen during micro-bursts, where the client typically is unaware that it should stop sending more data, while it is able to.

• connections

  The number of connect requests handled. This is an ever increasing number just counting how many connections were accepted.

• disconnects

  The number of disconnected clients. A disconnect either happens because the client goes away, or due to an idle timeout in the relay. The difference between this metric and connections is the amount of connections actively held by the relay. In normal situations this amount remains within reasonable bounds. Many connections, but few disconnections typically indicate a possible connection leak in the client. The idle connections disconnect in the relay here is to guard against resource drain in such scenarios.

• dispatch_wallTime_us

  The number of microseconds spent by the dispatchers to do their work. In particular on multi-core systems, this value can be confusing, however, it indicates how long the dispatchers were doing work handling clients. It includes everything they do, from reading data from a socket, cleaning up the input metric, to adding the metric to the appropriate queues. The larger the configuration, and more complex in terms of matches, the more time the dispatchers will spend on the cpu. But also time they do /not/ spend on the cpu is included in this number. It is the pure wallclock time the dispatcher was serving a client.

• dispatch_sleepTime_us

  The number of microseconds spent by the dispatchers sleeping waiting for work. When this value gets small (or even zero) the dispatcher has so much work that it doesn't sleep any more, and likely can't process the work in a timely fashion any more. This value plus the wallTime from above sort of sums up to the total uptime taken by this dispatcher. Therefore, expressing the wallTime as percentage of this sum gives the busyness percentage draining all the way up to 100% if sleepTime goes to 0.

• server_wallTime_us

  The number of microseconds spent by the servers to send the metrics from their queues. This value includes connection creation, reading from the queue, and sending metrics over the network.

• dispatcherX

  For each indivual dispatcher, the metrics received and blackholed plus the wall clock time. The values are as described above.

• destinations.X

  For all known destinations, the number of dropped, queued and sent metrics plus the wall clock time spent. The values are as described above.

• aggregators.metricsReceived

  The number of metrics that were matched an aggregator rule and were accepted by the aggregator. When a metric matches multiple aggregators, this value will reflect that. A metric is not counted when it is considered syntactically invalid, e.g. no value was found.

• aggregators.metricsDropped

  The number of metrics that were sent to an aggregator, but did not fit timewise. This is either because the metric was too far in the past or future. The expire after clause in aggregate statements controls how long in the past metric values are accepted.

• aggregators.metricsSent

  The number of metrics that were sent from the aggregators. These metrics were produced and are the actual results of aggregations.

The original argument for building carbon-c-relay was speed, with configurablility following close. To date, performance has bypassed the original carbon-relay.py by orders of magnitude, but the actual speed highly depends on perception and scenario. What follows below are some rough numbers about the environment at Booking.com where carbon-c-relay is used extensively in production.

carbon-c-relay runs on all of our machines as a local submission relay. Its config is simply a match all to a any_of cluster with a number of upstream relays to try and send the metrics to. These relays run with 4 workers, and receive a minimal amount of metrics per minute, typically between 50 and 200. These instances take typically around 19MiB of RAM and consume at top 0.8% CPU of a 2.4GHz core. The minimal footprint of the relay is a desired property for running on all of our machines.

The main relays we run, have roughly 20 clusters defined with fnv1a_ch hash. Average clustersize around 10 members. On top of that 30 match rules are defined. For a mildly-loaded relay receiving 1M metrics per minute, the relay consumes 750MiB of RAM and needs around 40% of a 2.4GHz core. A relay with more load but the same configuration, 3M metrics per minute, needs almost 2GiB of RAM, and some 45% CPU of a 2.4GHz core. The memory usage is mainly in the buffers for writing to the server stores.

On the stores, we run relays with a simple config with a match all rule to an any_of cluster pointing to 13 locally running carbon-cache.py instances. These relays receive up to 1.7M metrics per minute, and require some 110MiB RAM for that. The CPU usage is around 15% of a 2.4GHz core.

For aggregations we don't do much traffic (55K per minute) on a couple of aggregations expanding to a thousand of metrics. In our setup this takes 30MiB of RAM usage with some 30% CPU usage.

Fabian Groffen

This program was originally developed for Booking.com. With approval from Booking.com, the code was generalised and published as Open Source on github, for which the author would like to express his gratitude.