Merge pull request #4507 from apache/prometheus_metrics

feat: additional prometheus metrics

.devcontainer/devcontainer.json

@@ -26,4 +26,4 @@
     "postCreateCommand": "./configure && make",
 
     "extensions": ["erlang-ls.erlang-ls"]
-}
+}

src/couch/src/couch.app.src

@@ -47,7 +47,6 @@
         ioq,
         couch_stats,
         hyper,
-        couch_prometheus,
         couch_dist
     ]},
     {env, [

src/couch_prometheus/src/couch_prometheus.app.src

@@ -14,7 +14,7 @@
     {description, "Aggregated metrics info for Prometheus consumption"},
     {vsn, git},
     {registered, []},
-    {applications, [kernel, stdlib, folsom, couch_stats, couch_log]},
+    {applications, [kernel, stdlib, folsom, couch_stats, couch_log, mem3, couch]},
     {mod, {couch_prometheus_app, []}},
     {env, []}
 ]}.

src/couch_prometheus/src/couch_prometheus_server.erl

@@ -112,12 +112,33 @@ get_system_stats() ->
         get_message_queue_stats(),
         get_run_queue_stats(),
         get_vm_stats(),
-        get_ets_stats()
+        get_ets_stats(),
+        get_internal_replication_jobs_stat(),
+        get_membership_stat()
     ]).
 
 get_uptime_stat() ->
     to_prom(uptime_seconds, counter, "couchdb uptime", couch_app:uptime() div 1000).
 
+get_internal_replication_jobs_stat() ->
+    to_prom(
+        internal_replication_jobs,
+        gauge,
+        "count of internal replication changes to process",
+        mem3_sync:get_backlog()
+    ).
+
+get_membership_stat() ->
+    % expected nodes
+    ClusterNodes = mem3:nodes(),
+    % connected nodes
+    AllNodes = nodes([this, visible]),
+    Labels = [
+        {[{nodes, "cluster_nodes"}], length(ClusterNodes)},
+        {[{nodes, "all_nodes"}], length(AllNodes)}
+    ],
+    to_prom(membership, gauge, "count of nodes in the cluster", Labels).
+
 get_vm_stats() ->
     MemLabels = lists:map(
         fun({Type, Value}) ->
@@ -177,24 +198,27 @@ get_io_stats() ->
     ].
 
 get_message_queue_stats() ->
-    QLenFun = fun(Name) -> message_queue_len(whereis(Name)) end,
-    Queues = lists:map(QLenFun, registered()),
+    QFun = fun(Name) -> {Name, message_queue_len(whereis(Name))} end,
+    Queues = lists:map(QFun, registered()),
+    QueueLens = lists:map(fun({_, Len}) -> Len end, Queues),
+    QueueLenByLabel = lists:map(fun({Name, Len}) -> {[{queue_name, Name}], Len} end, Queues),
     [
         to_prom(
-            erlang_message_queues, gauge, "total size of all message queues", lists:sum(Queues)
+            erlang_message_queues, gauge, "total size of all message queues", lists:sum(QueueLens)
         ),
         to_prom(
             erlang_message_queue_min,
             gauge,
             "minimum size across all message queues",
-            lists:min(Queues)
+            lists:min(QueueLens)
         ),
         to_prom(
             erlang_message_queue_max,
             gauge,
             "maximum size across all message queues",
-            lists:max(Queues)
-        )
+            lists:max(QueueLens)
+        ),
+        to_prom(erlang_message_queue_size, gauge, "size of message queue", QueueLenByLabel)
     ].
 
 message_queue_len(undefined) ->
@@ -247,19 +271,6 @@ update_refresh_timer() ->
 
 -include_lib("couch/include/couch_eunit.hrl").
 
-system_stats_test() ->
-    lists:foreach(
-        fun(Line) ->
-            ?assert(is_binary(Line)),
-            Trimmed = string:trim(Line),
-            ?assert(starts_with(<<"couchdb_">>, Trimmed) orelse starts_with(<<"# ">>, Trimmed))
-        end,
-        get_system_stats()
-    ).
-
-starts_with(Prefix, Line) when is_binary(Prefix), is_binary(Line) ->
-    binary:longest_common_prefix([Prefix, Line]) > 0.
-
 message_queue_len_test() ->
     self() ! refresh,
     ?assert(message_queue_len(self()) >= 1),

src/couch_prometheus/src/couch_prometheus_util.erl

@@ -71,6 +71,19 @@ couch_to_prom([couchdb, httpd_status_codes, 200], Info, _All) ->
     });
 couch_to_prom([couchdb, httpd_status_codes, Code], Info, _All) ->
     to_prom(httpd_status_codes, {[{code, Code}], val(Info)});
+% Convert to gauge in prometheus type. This is required because
+% prometheus assumes that counters are cumulative and should be
+% rated by default, whereas folsom (the library CouchDB uses for
+% metrics) allows counters to be decremented as well. Folsom supports
+% gauges but does not track their state to allow increment/decrement.
+% Basically, anywhere we use couch_stats:decrement_count we should
+% be converting to a prometheus gauge.
+couch_to_prom([couchdb, open_databases], Info, _All) ->
+    to_prom(open_databases, gauge, desc(Info), val(Info));
+couch_to_prom([couchdb, open_os_files], Info, _All) ->
+    to_prom(open_os_files, gauge, desc(Info), val(Info));
+couch_to_prom([couchdb, httpd, clients_requesting_changes], Info, _All) ->
+    to_prom(httpd_clients_requesting_changes, gauge, desc(Info), val(Info));
 couch_to_prom([ddoc_cache, hit], Info, All) ->
     Total = val(Info) + val([ddoc_cache, miss], All),
     to_prom(ddoc_cache_requests_total, counter, "number of design doc cache requests", Total);
@@ -109,9 +122,13 @@ type_def(Metric, Type, Desc) ->
         to_bin(io_lib:format("# TYPE ~s ~s", [Name, Type]))
     ].
 
-to_prom(Metric, Type, Desc, Data) ->
+% support creating a metric series with multiple label/values.
+% Instances is of the form [{[{LabelName, LabelValue}], Value}, ...]
+to_prom(Metric, Type, Desc, Instances) when is_list(Instances) ->
     TypeStr = type_def(Metric, Type, Desc),
-    [TypeStr] ++ to_prom(Metric, Data).
+    [TypeStr] ++ lists:flatmap(fun(Inst) -> to_prom(Metric, Inst) end, Instances);
+to_prom(Metric, Type, Desc, Data) ->
+    to_prom(Metric, Type, Desc, [Data]).
 
 to_prom(Metric, Instances) when is_list(Instances) ->
     lists:flatmap(fun(Inst) -> to_prom(Metric, Inst) end, Instances);

src/couch_prometheus/test/eunit/couch_prometheus_e2e_tests.erl

@@ -40,7 +40,8 @@ e2e_test_() ->
                     ?TDEF_FE(t_chttpd_port),
                     ?TDEF_FE(t_prometheus_port),
                     ?TDEF_FE(t_metric_updated),
-                    ?TDEF_FE(t_no_duplicate_metrics)
+                    ?TDEF_FE(t_no_duplicate_metrics),
+                    ?TDEF_FE(t_starts_with_couchdb)
                 ]
             }
         }
@@ -70,7 +71,7 @@ reject_test_() ->
     }.
 
 setup_prometheus(WithAdditionalPort) ->
-    Ctx = test_util:start_couch([chttpd]),
+    Ctx = test_util:start_couch([mem3, chttpd, couch_prometheus]),
     Persist = false,
     Hashed = couch_passwords:hash_admin_password(?PASS),
     ok = config:set("admins", ?USER, binary_to_list(Hashed), Persist),
@@ -145,6 +146,33 @@ t_metric_updated(Port) ->
         end
     ).
 
+t_starts_with_couchdb(Port) ->
+    Url = node_local_url(Port),
+    Stats = get_stats(Url),
+    Lines = re:split(Stats, "\n"),
+    lists:foreach(
+        fun(Line) ->
+            ?assert(is_binary(Line)),
+            Trimmed = string:trim(Line),
+            Expect = "^(#|couchdb_|$)",
+            case re:run(Trimmed, Expect) of
+                {match, _} ->
+                    ok;
+                nomatch ->
+                    erlang:error(
+                        {assertRegexp_failed, [
+                            {module, ?MODULE},
+                            {line, ?LINE},
+                            {regexp, (Trimmed)},
+                            {expected_to_match, Expect},
+                            {result, nomatch}
+                        ]}
+                    )
+            end
+        end,
+        Lines
+    ).
+
 node_local_url(Port) ->
     Addr = config:get("chttpd", "bind_address", "127.0.0.1"),
     lists:concat(["http://", Addr, ":", Port, "/_node/_local/_prometheus"]).