{ "componentChunkName": "component---src-templates-issues-tsx", "path": "/issues/453", "result": {"data":{"issuesJson":{"id":"bd898015-ea7e-5b60-9ada-89c9d3627f9d","title":"ncabatoff/process-exporter: 用于进程指标的 Prometheus Exporter","number":453,"bodyHTML":"
\n



\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Reposncabatoff/process-exporter
Views\"views\"
Stars\"stars\"
Forks\"forks\"
License\"license\"
UpdatedAt\"last-commit\"
CreatedAt\"create-at\"
\n\n

\"\"

\n\n
\n

process-exporter

\n

Prometheus exporter that mines /proc to report on selected processes.

\n

\"Release\"
\n\"Powered
\n\"Build\"

\n

Some apps are impractical to instrument directly, either because you
\ndon't control the code or they're written in a language that isn't easy to
\ninstrument with Prometheus. We must instead resort to mining /proc.

\n

Installation

\n

Either grab a package for your OS from the Releases page, or
\ninstall via docker.

\n

Running

\n

Usage:

\n
  process-exporter [options] -config.path filename.yml\n
\n

or via docker:

\n
  docker run -d --rm -p 9256:9256 --privileged -v /proc:/host/proc -v `pwd`:/config ncabatoff/process-exporter --procfs /host/proc -config.path /config/filename.yml\n\n
\n

Important options (run process-exporter --help for full list):

\n

-children (default:true) makes it so that any process that otherwise
\nisn't part of its own group becomes part of the first group found (if any) when
\nwalking the process tree upwards. In other words, resource usage of
\nsubprocesses is added to their parent's usage unless the subprocess identifies
\nas a different group name.

\n

-threads (default:true) means that metrics will be broken down by thread name
\nas well as group name.

\n

-recheck (default:false) means that on each scrape the process names are
\nre-evaluated. This is disabled by default as an optimization, but since
\nprocesses can choose to change their names, this may result in a process
\nfalling into the wrong group if we happen to see it for the first time before
\nit's assumed its proper name. You can use -recheck-with-time-limit to enable this
\nfeature only for a specific duration after process starts.

\n

-procnames is intended as a quick alternative to using a config file. Details
\nin the following section.

\n

-remove-empty-groups (default:false) forget process groups with no processes.
\nThis is particularly useful if you have some process groups that you expect will
\nnever return (e.g. if you have process groups named \"scan-\", and once
\nthe scan is completed no more process will ever run for that scan again).

\n

To disable any of these options, use the -option=false.

\n

Configuration and group naming

\n

To select and group the processes to monitor, either provide command-line
\narguments or use a YAML configuration file.

\n

The recommended option is to use a config file via -config.path, but for
\nconvenience and backwards compatibility the -procnames/-namemapping options
\nexist as an alternative.

\n

Using a config file

\n

The general format of the -config.path YAML file is a top-level
\nprocess_names section, containing a list of name matchers:

\n
process_names:\n  - matcher1\n  - matcher2\n  ...\n  - matcherN\n
\n

The default config shipped with the deb/rpm packages is:

\n
process_names:\n  - name: \"{{.Comm}}\"\n    cmdline:\n    - '.+'\n
\n

A process may only belong to one group: even if multiple items would match, the
\nfirst one listed in the file wins.

\n

(Side note: to avoid confusion with the cmdline YAML element, we'll refer to
\nthe command-line arguments of a process /proc/<pid>/cmdline as the array
\nargv[].)

\n

Using a config file: group name

\n

Each item in process_names gives a recipe for identifying and naming
\nprocesses. The optional name tag defines a template to use to name
\nmatching processes; if not specified, name defaults to {{.ExeBase}}.

\n

Template variables available:

\n\n

Using PID or StartTime is discouraged: this is almost never what you want,
\nand is likely to result in high cardinality metrics which Prometheus will have
\ntrouble with.

\n

Using a config file: process selectors

\n

Each item in process_names must contain one or more selectors (comm, exe
\nor cmdline); if more than one selector is present, they must all match. Each
\nselector is a list of strings to match against a process's comm, argv[0],
\nor in the case of cmdline, a regexp to apply to the command line. The cmdline
\nregexp uses the Go syntax.

\n

For comm and exe, the list of strings is an OR, meaning any process
\nmatching any of the strings will be added to the item's group.

\n

For cmdline, the list of regexes is an AND, meaning they all must match. Any
\ncapturing groups in a regexp must use the ?P<name> option to assign a name to
\nthe capture, which is used to populate .Matches.

\n

Performance tip: give an exe or comm clause in addition to any cmdline
\nclause, so you avoid executing the regexp when the executable name doesn't
\nmatch.

\n
\nprocess_names:\n  # comm is the second field of /proc/<pid>/stat minus parens.\n  # It is the base executable name, truncated at 15 chars.\n  # It cannot be modified by the program, unlike exe.\n  - comm:\n    - bash\n\n  # exe is argv[0]. If no slashes, only basename of argv[0] need match.\n  # If exe contains slashes, argv[0] must match exactly.\n  - exe:\n    - postgres\n    - /usr/local/bin/prometheus\n\n  # cmdline is a list of regexps applied to argv.\n  # Each must match, and any captures are added to the .Matches map.\n  - name: \"{{.ExeFull}}:{{.Matches.Cfgfile}}\"\n    exe:\n    - /usr/local/bin/process-exporter\n    cmdline:\n    - -config.path\\s+(?P<Cfgfile>\\S+)\n\n
\n

Here's the config I use on my home machine:

\n
\nprocess_names:\n  - comm:\n    - chromium-browse\n    - bash\n    - prometheus\n    - gvim\n  - exe:\n    - /sbin/upstart\n    cmdline:\n    - --user\n    name: upstart:-user\n\n
\n

Using -procnames/-namemapping instead of config.path

\n

Every name in the procnames list becomes a process group. The default name of
\na process is the value found in the second field of /proc//stat
\n(\"comm\"), which is truncated at 15 chars. Usually this is the same as the
\nname of the executable.

\n

If -namemapping isn't provided, every process with a comm value present
\nin -procnames is assigned to a group based on that name, and any other
\nprocesses are ignored.

\n

The -namemapping option is a comma-separated list of alternating
\nname,regexp values. It allows assigning a name to a process based on a
\ncombination of the process name and command line. For example, using

\n

-namemapping \"python2,([^/]+).py,java,-jar\\s+([^/]+).jar\"

\n

will make it so that each different python2 and java -jar invocation will be
\ntracked with distinct metrics. Processes whose remapped name is absent from
\nthe procnames list will be ignored. On a Ubuntu Xenian machine being used as
\na workstation, here's a good way of tracking resource usage for a few
\ndifferent key user apps:

\n

process-exporter -namemapping \"upstart,(--user)\"
\n-procnames chromium-browse,bash,gvim,prometheus,process-exporter,upstart:-user

\n

Since upstart --user is the parent process of the X11 session, this will
\nmake all apps started by the user fall into the group named \"upstart:-user\",
\nunless they're one of the others named explicitly with -procnames, like gvim.

\n

Group Metrics

\n

There's no meaningful way to name a process that will only ever name a single process, so process-exporter assumes that every metric will be attached
\nto a group of processes - not a
\nprocess group in the technical
\nsense, just one or more processes that meet a configuration's specification
\nof what should be monitored and how to name it.

\n

All these metrics start with namedprocess_namegroup_ and have at minimum
\nthe label groupname.

\n

num_procs gauge

\n

Number of processes in this group.

\n

cpu_seconds_total counter

\n

CPU usage based on /proc/[pid]/stat fields utime(14) and stime(15) i.e. user and system time. This is similar to the node_exporter's node_cpu_seconds_total.

\n

read_bytes_total counter

\n

Bytes read based on /proc/[pid]/io field read_bytes. The man page
\nsays

\n
\n

Attempt to count the number of bytes which this process really did cause to be fetched from the storage layer. This is accurate for block-backed filesystems.

\n
\n

but I would take it with a grain of salt.

\n

As /proc/[pid]/io are set by the kernel as read only to the process' user (see #137), to get these values you should run process-exporter either as that user or as root. Otherwise, we can't read these values and you'll get a constant 0 in the metric.

\n

write_bytes_total counter

\n

Bytes written based on /proc/[pid]/io field write_bytes. As with
\nread_bytes, somewhat dubious. May be useful for isolating which processes
\nare doing the most I/O, but probably not measuring just how much I/O is happening.

\n

major_page_faults_total counter

\n

Number of major page faults based on /proc/[pid]/stat field majflt(12).

\n

minor_page_faults_total counter

\n

Number of minor page faults based on /proc/[pid]/stat field minflt(10).

\n

context_switches_total counter

\n

Number of context switches based on /proc/[pid]/status fields voluntary_ctxt_switches
\nand nonvoluntary_ctxt_switches. The extra label ctxswitchtype can have two values:
\nvoluntary and nonvoluntary.

\n

memory_bytes gauge

\n

Number of bytes of memory used. The extra label memtype can have three values:

\n

resident: Field rss(24) from /proc/[pid]/stat, whose doc says:

\n
\n

This is just the pages which count toward text, data, or stack space. This does not include pages which have not been demand-loaded in, or which are swapped out.

\n
\n

virtual: Field vsize(23) from /proc/[pid]/stat, virtual memory size.

\n

swapped: Field VmSwap from /proc/[pid]/status, translated from KB to bytes.

\n

If gathering smaps file is enabled, two additional values for memtype are added:

\n

proportionalResident: Sum of \"Pss\" fields from /proc/[pid]/smaps, whose doc says:

\n
\n

The \"proportional set size\" (PSS) of a process is the count of pages it has
\nin memory, where each page is divided by the number of processes sharing it.

\n
\n

proportionalSwapped: Sum of \"SwapPss\" fields from /proc/[pid]/smaps

\n

open_filedesc gauge

\n

Number of file descriptors, based on counting how many entries are in the directory
\n/proc/[pid]/fd.

\n

worst_fd_ratio gauge

\n

Worst ratio of open filedescs to filedesc limit, amongst all the procs in the
\ngroup. The limit is the fd soft limit based on /proc/[pid]/limits.

\n

Normally Prometheus metrics ought to be as \"basic\" as possible (i.e. the raw
\nvalues rather than a derived ratio), but we use a ratio here because nothing
\nelse makes sense. Suppose there are 10 procs in a given group, each with a
\nsoft limit of 4096, and one of them has 4000 open fds and the others all have
\n40, their total fdcount is 4360 and total soft limit is 40960, so the ratio
\nis 1:10, but in fact one of the procs is about to run out of fds. With
\nworst_fd_ratio we're able to know this: in the above example it would be
\n0.97, rather than the 0.10 you'd see if you computed sum(open_filedesc) /
\nsum(limit_filedesc).

\n

oldest_start_time_seconds gauge

\n

Epoch time (seconds since 1970/1/1) at which the oldest process in the group
\nstarted. This is derived from field starttime(22) from /proc/[pid]/stat, added
\nto boot time to make it relative to epoch.

\n

num_threads gauge

\n

Sum of number of threads of all process in the group. Based on field num_threads(20)
\nfrom /proc/[pid]/stat.

\n

states gauge

\n

Number of threads in the group in each of various states, based on the field
\nstate(3) from /proc/[pid]/stat.

\n

The extra label state can have these values: Running, Sleeping, Waiting, Zombie, Other.

\n

Group Thread Metrics

\n

Since publishing thread metrics adds a lot of overhead, use the -threads command-line argument to disable them,
\nif necessary.

\n

All these metrics start with namedprocess_namegroup_ and have at minimum
\nthe labels groupname and threadname. threadname is field comm(2) from
\n/proc/[pid]/stat. Just as groupname breaks the set of processes down into
\ngroups, threadname breaks a given process group down into subgroups.

\n

thread_count gauge

\n

Number of threads in this thread subgroup.

\n

thread_cpu_seconds_total counter

\n

Same as cpu_user_seconds_total and cpu_system_seconds_total, but broken down
\nper-thread subgroup. Unlike cpu_user_seconds_total/cpu_system_seconds_total,
\nthe label cpumode is used to distinguish between user and system time.

\n

thread_io_bytes_total counter

\n

Same as read_bytes_total and write_bytes_total, but broken down
\nper-thread subgroup. Unlike read_bytes_total/write_bytes_total,
\nthe label iomode is used to distinguish between read and write bytes.

\n

thread_major_page_faults_total counter

\n

Same as major_page_faults_total, but broken down per-thread subgroup.

\n

thread_minor_page_faults_total counter

\n

Same as minor_page_faults_total, but broken down per-thread subgroup.

\n

thread_context_switches_total counter

\n

Same as context_switches_total, but broken down per-thread subgroup.

\n

Instrumentation cost

\n

process-exporter will consume CPU in proportion to the number of processes in
\nthe system and the rate at which new ones are created. The most expensive
\nparts - applying regexps and executing templates - are only applied once per
\nprocess seen, unless the command-line option -recheck is provided.

\n

If you have mostly long-running processes process-exporter overhead should be
\nminimal: each time a scrape occurs, it will parse of /proc/$pid/stat and
\n/proc/$pid/cmdline for every process being monitored and add a few numbers.

\n

Dashboards

\n

An example Grafana dashboard to view the metrics is available at https://grafana.net/dashboards/249

\n

Building

\n

Requires Go 1.21 (at least) installed.

\n
make\n
\n

Exposing metrics through HTTPS

\n

web-config.yml

\n
# Minimal TLS configuration example. Additionally, a certificate and a key file\n# are needed.\ntls_server_config:\n  cert_file: server.crt\n  key_file: server.key\n
\n

Running

\n
$ ./process-exporter -web.config.file web-config.yml &\n$ curl -sk https://localhost:9256/metrics | grep process\n\n# HELP namedprocess_scrape_errors general scrape errors: no proc metrics collected during a cycle\n# TYPE namedprocess_scrape_errors counter\nnamedprocess_scrape_errors 0\n# HELP namedprocess_scrape_partial_errors incremented each time a tracked proc's metrics collection fails partially, e.g. unreadable I/O stats\n# TYPE namedprocess_scrape_partial_errors counter\nnamedprocess_scrape_partial_errors 0\n# HELP namedprocess_scrape_procread_errors incremented each time a proc's metrics collection fails\n# TYPE namedprocess_scrape_procread_errors counter\nnamedprocess_scrape_procread_errors 0\n# HELP process_cpu_seconds_total Total user and system CPU time spent in seconds.\n# TYPE process_cpu_seconds_total counter\nprocess_cpu_seconds_total 0.21\n# HELP process_exporter_build_info A metric with a constant '1' value labeled by version, revision, branch, and goversion from which process_exporter was built.\n# TYPE process_exporter_build_info gauge\nprocess_exporter_build_info{branch=\"\",goversion=\"go1.17.3\",revision=\"\",version=\"\"} 1\n# HELP process_max_fds Maximum number of open file descriptors.\n# TYPE process_max_fds gauge\nprocess_max_fds 1.048576e+06\n# HELP process_open_fds Number of open file descriptors.\n# TYPE process_open_fds gauge\nprocess_open_fds 10\n
\n

For further information about TLS configuration, please visit: exporter-toolkit

","updatedAt":"2025-05-21T16:16:53Z","upvoteCount":null,"author":{"login":"eryajf","avatarUrl":"https://avatars.githubusercontent.com/u/33259379?u=e4a4090a38ac2473aaed4ef9945233636776c6c3&v=4","url":"https://github.com/eryajf"},"category":null,"labels":{"edges":[{"node":{"name":"更多","color":"25B472"}},{"node":{"name":"Prometheus-Exporter","color":"367acb"}},{"node":{"name":"ncabatoff","color":"4092f5"}}]},"comments":{"edges":[]}}},"pageContext":{"number":453,"previous":{"title":"denilsonsa/prettyping: prettyping 是标准工具的包装器 ping，目的是使输出更漂亮、更丰富多彩、更紧凑且更易于阅读。","number":452},"next":{"title":"dalance/procs: 用 Rust 编写的 ps 的现代替代品","number":454}}}, "staticQueryHashes": ["151096407","2861350382"]}