Statistics: Rewrite statistics backends to use OpenMetrics
authorJonas Schäfer <jonas@wielicki.name>
Sun, 18 Apr 2021 11:47:41 +0200
changeset 11527 5f15ab7c6ae5
parent 11526 5bd38d9197e1
child 11528 6de302b53a3e
Statistics: Rewrite statistics backends to use OpenMetrics The metric subsystem of Prosody has had some shortcomings from the perspective of the current state-of-the-art in metric observability. The OpenMetrics standard [0] is a formalization of the data model (and serialization format) of the well-known and widely-used Prometheus [1] software stack. The previous stats subsystem of Prosody did not map well to that format (see e.g. [2] and [3]); the key reason is that it was trying to do too much math on its own ([2]) while lacking first-class support for "families" of metrics ([3]) and structured metric metadata (despite the `extra` argument to metrics, there was no standard way of representing common things like "tags" or "labels"). Even though OpenMetrics has grown from the Prometheus world of monitoring, it maps well to other popular monitoring stacks such as: - InfluxDB (labels can be mapped to tags and fields as necessary) - Carbon/Graphite (labels can be attached to the metric name with dot-separation) - StatsD (see graphite when assuming that graphite is used as backend, which is the default) The util.statsd module has been ported to use the OpenMetrics model as a proof of concept. An implementation which exposes the util.statistics backend data as Prometheus metrics is ready for publishing in prosody-modules (most likely as mod_openmetrics_prometheus to avoid breaking existing 0.11 deployments). At the same time, the previous measure()-based API had one major advantage: It is really simple and easy to use without requiring lots of knowledge about OpenMetrics or similar concepts. For that reason as well as compatibility with existing code, it is preserved and may even be extended in the future. However, code relying on the `stats-updated` event as well as `get_stats` from `statsmanager` will break because the data model has changed completely; in case of `stats-updated`, the code will simply not run (as the event was renamed in order to avoid conflicts); the `get_stats` function has been removed completely (so it will cause a traceback when it is attempted to be used). Note that the measure_*_event methods have been removed from the module API. I was unable to find any uses or documentation and thus deemed they should not be ported. Re-implementation is possible when necessary. [0]: https://openmetrics.io/ [1]: https://prometheus.io/ [2]: #959 [3]: #960
.luacheckrc
core/moduleapi.lua
core/statsmanager.lua
plugins/mod_admin_shell.lua
util/openmetrics.lua
util/statistics.lua
util/statsd.lua
--- a/.luacheckrc	Sun Apr 25 17:32:27 2021 +0200
+++ b/.luacheckrc	Sun Apr 18 11:47:41 2021 +0200
@@ -80,9 +80,7 @@
 		"module.log",
 		"module.log_status",
 		"module.measure",
-		"module.measure_event",
-		"module.measure_global_event",
-		"module.measure_object_event",
+		"module.metric",
 		"module.open_store",
 		"module.provides",
 		"module.remove_item",
--- a/core/moduleapi.lua	Sun Apr 25 17:32:27 2021 +0200
+++ b/core/moduleapi.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -510,26 +510,33 @@
 
 function api:measure(name, stat_type, conf)
 	local measure = require "core.statsmanager".measure;
-	return measure(stat_type, "/"..self.host.."/mod_"..self.name.."/"..name, conf);
+	local fixed_label_key, fixed_label_value
+	if self.host ~= "*" then
+		fixed_label_key = "host"
+		fixed_label_value = self.host
+	end
+	-- new_legacy_metric takes care of scoping for us, as it does not accept
+	-- an array of labels
+	-- the prosody_ prefix is automatically added by statsmanager for legacy
+	-- metrics.
+	return measure(stat_type, "mod_"..self.name.."/"..name, conf, fixed_label_key, fixed_label_value)
 end
 
-function api:measure_object_event(events_object, event_name, stat_name)
-	local m = self:measure(stat_name or event_name, "times");
-	local function handler(handlers, _event_name, _event_data)
-		local finished = m();
-		local ret = handlers(_event_name, _event_data);
-		finished();
-		return ret;
+function api:metric(type_, name, unit, description, label_keys, conf)
+	local metric = require "core.statsmanager".metric;
+	local is_scoped = self.host ~= "*"
+	if is_scoped then
+		-- prepend `host` label to label keys if this is not a global module
+		local orig_labels = label_keys
+		label_keys = array { "host" }
+		label_keys:append(orig_labels)
 	end
-	return self:hook_object_event(events_object, event_name, handler);
-end
-
-function api:measure_event(event_name, stat_name)
-	return self:measure_object_event((hosts[self.host] or prosody).events.wrappers, event_name, stat_name);
-end
-
-function api:measure_global_event(event_name, stat_name)
-	return self:measure_object_event(prosody.events.wrappers, event_name, stat_name);
+	local mf = metric(type_, "prosody_mod_"..self.name.."/"..name, unit, description, label_keys, conf)
+	if is_scoped then
+		-- make sure to scope the returned metric family to the current host
+		return mf:with_partial_label(self.host)
+	end
+	return mf
 end
 
 local status_priorities = { error = 3, warn = 2, info = 1, core = 0 };
--- a/core/statsmanager.lua	Sun Apr 25 17:32:27 2021 +0200
+++ b/core/statsmanager.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -3,6 +3,8 @@
 local log = require "util.logger".init("stats");
 local timer = require "util.timer";
 local fire_event = prosody.events.fire_event;
+local array = require "util.array";
+local timed = require "util.openmetrics".timed;
 
 local stats_interval_config = config.get("*", "statistics_interval");
 local stats_interval = tonumber(stats_interval_config);
@@ -57,15 +59,149 @@
 	log("error", "Error loading statistics provider '%s': %s", stats_provider, stats_err);
 end
 
-local measure, collect;
-local latest_stats = {};
-local changed_stats = {};
-local stats_extra = {};
+local measure, collect, metric, cork, uncork;
 
 if stats then
-	function measure(type, name, conf)
-		local f = assert(stats[type], "unknown stat type: "..type);
-		return f(name, conf);
+	function metric(type_, name, unit, description, labels, extra)
+		local registry = stats.metric_registry
+		local f = assert(registry[type_], "unknown metric family type: "..type_);
+		return f(registry, name, unit or "", description or "", labels, extra);
+	end
+
+	local function new_legacy_metric(stat_type, name, unit, description, fixed_label_key, fixed_label_value, extra)
+		local label_keys = array()
+		local conf = extra or {}
+		if fixed_label_key then
+			label_keys:push(fixed_label_key)
+		end
+		unit = unit or ""
+		local mf = metric(stat_type, "prosody_" .. name, unit, description, label_keys, conf);
+		if fixed_label_key then
+			mf = mf:with_partial_label(fixed_label_value)
+		end
+		return mf:with_labels()
+	end
+
+	local function unwrap_legacy_extra(extra, type_, name, unit)
+		local description = extra and extra.description or "Legacy "..type_.." metric "..name
+		unit = extra and extra.unit or unit
+		return description, unit
+	end
+
+	-- These wrappers provide the pre-OpenMetrics interface of statsmanager
+	-- and moduleapi (module:measure).
+	local legacy_metric_wrappers = {
+		amount = function(name, fixed_label_key, fixed_label_value, extra)
+			local initial = 0
+			if type(extra) == "number" then
+				initial = extra
+			else
+				initial = extra and extra.initial or initial
+			end
+			local description, unit = unwrap_legacy_extra(extra, "amount", name)
+
+			local m = new_legacy_metric("gauge", name, unit, description, fixed_label_key, fixed_label_value)
+			m:set(initial or 0)
+			return function(v)
+				m:set(v)
+			end
+		end;
+
+		counter = function(name, fixed_label_key, fixed_label_value, extra)
+			if type(extra) == "number" then
+				-- previous versions of the API allowed passing an initial
+				-- value here; we do not allow that anymore, it is not a thing
+				-- which makes sense with counters
+				extra = nil
+			end
+
+			local description, unit = unwrap_legacy_extra(extra, "counter", name)
+
+			local m = new_legacy_metric("counter", name, unit, description, fixed_label_key, fixed_label_value)
+			m:set(0)
+			return function(v)
+				m:add(v)
+			end
+		end;
+
+		rate = function(name, fixed_label_key, fixed_label_value, extra)
+			if type(extra) == "number" then
+				-- previous versions of the API allowed passing an initial
+				-- value here; we do not allow that anymore, it is not a thing
+				-- which makes sense with counters
+				extra = nil
+			end
+
+			local description, unit = unwrap_legacy_extra(extra, "counter", name)
+
+			local m = new_legacy_metric("counter", name, unit, description, fixed_label_key, fixed_label_value)
+			m:set(0)
+			return function()
+				m:add(1)
+			end
+		end;
+
+		times = function(name, fixed_label_key, fixed_label_value, extra)
+			local conf = {}
+			if extra and extra.buckets then
+				conf.buckets = extra.buckets
+			else
+				conf.buckets = { 0.001, 0.01, 0.1, 1.0, 10.0, 100.0 }
+			end
+			local description, _ = unwrap_legacy_extra(extra, "times", name)
+
+			local m = new_legacy_metric("histogram", name, "seconds", description, fixed_label_key, fixed_label_value, conf)
+			return function()
+				return timed(m)
+			end
+		end;
+
+		sizes = function(name, fixed_label_key, fixed_label_value, extra)
+			local conf = {}
+			if extra and extra.buckets then
+				conf.buckets = extra.buckets
+			else
+				conf.buckets = { 1024, 4096, 32768, 131072, 1048576, 4194304, 33554432, 134217728, 1073741824 }
+			end
+			local description, _ = unwrap_legacy_extra(extra, "sizes", name)
+
+			local m = new_legacy_metric("histogram", name, "bytes", description, fixed_label_key, fixed_label_value, conf)
+			return function(v)
+				m:sample(v)
+			end
+		end;
+
+		distribution = function(name, fixed_label_key, fixed_label_value, extra)
+			if type(extra) == "string" then
+				-- compat with previous API
+				extra = { unit = extra }
+			end
+			local description, unit = unwrap_legacy_extra(extra, "distribution", name, "")
+			local m = new_legacy_metric("summary", name, unit, description, fixed_label_key, fixed_label_value)
+			return function(v)
+				m:sample(v)
+			end
+		end;
+	};
+
+	-- Argument order switched here to support the legacy statsmanager.measure
+	-- interface.
+	function measure(stat_type, name, extra, fixed_label_key, fixed_label_value)
+		local wrapper = assert(legacy_metric_wrappers[stat_type], "unknown legacy metric type "..stat_type)
+		return wrapper(name, fixed_label_key, fixed_label_value, extra)
+	end
+
+	if stats.cork then
+		function cork()
+			return stats:cork()
+		end
+
+		function uncork()
+			return stats:uncork()
+		end
+	else
+		function cork() end
+		function uncork() end
 	end
 
 	if stats_interval or stats_interval_config == "manual" then
@@ -76,27 +212,23 @@
 		function collect()
 			local mark_collection_done = mark_collection_start();
 			fire_event("stats-update");
+			-- ensure that the backend is uncorked, in case it got stuck at
+			-- some point, to avoid infinite resource use
+			uncork()
 			mark_collection_done();
+			local manual_result = nil
 
-			if stats.get_stats then
+			if stats.metric_registry then
+				-- only if supported by the backend, we fire the event which
+				-- provides the current metric values
 				local mark_processing_done = mark_processing_start();
-				changed_stats, stats_extra = {}, {};
-				for stat_name, getter in pairs(stats.get_stats()) do
-					-- luacheck: ignore 211/type
-					local type, value, extra = getter();
-					local old_value = latest_stats[stat_name];
-					latest_stats[stat_name] = value;
-					if value ~= old_value then
-						changed_stats[stat_name] = value;
-					end
-					if extra then
-						stats_extra[stat_name] = extra;
-					end
-				end
-				fire_event("stats-updated", { stats = latest_stats, changed_stats = changed_stats, stats_extra = stats_extra });
+				local metric_registry = stats.metric_registry;
+				fire_event("openmetrics-updated", { metric_registry = metric_registry })
 				mark_processing_done();
+				manual_result = metric_registry;
 			end
-			return stats_interval;
+
+			return stats_interval, manual_result;
 		end
 		if stats_interval then
 			log("debug", "Statistics enabled using %s provider, collecting every %d seconds", stats_provider_name, stats_interval);
@@ -112,6 +244,22 @@
 else
 	log("debug", "Statistics disabled");
 	function measure() return measure; end
+
+	local dummy_mt = {}
+	function dummy_mt.__newindex()
+	end
+	function dummy_mt:__index()
+		return self
+	end
+	function dummy_mt:__call()
+		return self
+	end
+	local dummy = {}
+	setmetatable(dummy, dummy_mt)
+
+	function metric() return dummy; end
+	function cork() end
+	function uncork() end
 end
 
 local exported_collect = nil;
@@ -122,10 +270,10 @@
 return {
 	collect = exported_collect;
 	measure = measure;
-	get_stats = function ()
-		return latest_stats, changed_stats, stats_extra;
-	end;
-	get = function (name)
-		return latest_stats[name], stats_extra[name];
+	cork = cork;
+	uncork = uncork;
+	metric = metric;
+	get_metric_registry = function ()
+		return stats and stats.metric_registry or nil
 	end;
 };
--- a/plugins/mod_admin_shell.lua	Sun Apr 25 17:32:27 2021 +0200
+++ b/plugins/mod_admin_shell.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -36,6 +36,9 @@
 local serialize_config = serialization.new ({ fatal = false, unquoted = true});
 local time = require "util.time";
 
+local t_insert = table.insert;
+local t_concat = table.concat;
+
 local format_number = require "util.human.units".format;
 local format_table = require "util.human.io".table;
 
@@ -1342,187 +1345,112 @@
 	bytes = "B",
 };
 
-local function format_stat(type, unit, value, ref_value)
-	ref_value = ref_value or value;
-	--do return tostring(value) end
-	if not unit then
-		if type == "duration" then
-			unit = "seconds"
-		elseif type == "size" then
-			unit = "bytes";
-		elseif type == "rate" then
-			unit = " events/sec"
-			if ref_value < 0.9 then
-				unit = "events/min"
-				value = value*60;
-				if ref_value < 0.6/60 then
-					unit = "events/h"
-					value = value*60;
-				end
+local stats_methods = {};
+
+function stats_methods:render_single_fancy_histogram_ex(print, prefix, metric_family, metric, cumulative)
+	local creation_timestamp, sum, count
+	local buckets = {}
+	local prev_bucket_count = 0
+	for suffix, extra_labels, value in metric:iter_samples() do
+		if suffix == "_created" then
+			creation_timestamp = value
+		elseif suffix == "_sum" then
+			sum = value
+		elseif suffix == "_count" then
+			count = value
+		else
+			local bucket_threshold = extra_labels["le"]
+			local bucket_count
+			if cumulative then
+				bucket_count = value
+			else
+				bucket_count = value - prev_bucket_count
+				prev_bucket_count = value
 			end
-			return ("%.3g %s"):format(value, unit);
+			if bucket_threshold == "+Inf" then
+				t_insert(buckets, {threshold = 1/0, count = bucket_count})
+			elseif bucket_threshold ~= nil then
+				t_insert(buckets, {threshold = tonumber(bucket_threshold), count = bucket_count})
+			end
 		end
 	end
-	return format_number(value, short_units[unit] or unit or "", unit == "bytes" and 'b' or nil);
-end
+
+	if #buckets == 0 or not creation_timestamp or not sum or not count then
+		print("[no data or not a histogram]")
+		return false
+	end
 
-local stats_methods = {};
-function stats_methods:bounds(_lower, _upper)
-	for _, stat_info in ipairs(self) do
-		local data = stat_info[4];
-		if data then
-			local lower = _lower or data.min;
-			local upper = _upper or data.max;
-			local new_data = {
-				min = lower;
-				max = upper;
-				samples = {};
-				sample_count = 0;
-				count = data.count;
-				units = data.units;
-			};
-			local sum = 0;
-			for _, v in ipairs(data.samples) do
-				if v > upper then
-					break;
-				elseif v>=lower then
-					table.insert(new_data.samples, v);
-					sum = sum + v;
-				end
-			end
-			new_data.sample_count = #new_data.samples;
-			stat_info[4] = new_data;
-			stat_info[3] = sum/new_data.sample_count;
+	local graph_width, graph_height, wscale = #buckets, 10, 1;
+	if graph_width < 8 then
+		wscale = 8
+	elseif graph_width < 16 then
+		wscale = 4
+	elseif graph_width < 32 then
+		wscale = 2
+	end
+	local eighth_chars = "   ▁▂▃▄▅▆▇█";
+
+	local max_bin_samples = 0
+	for _, bucket in ipairs(buckets) do
+		if bucket.count > max_bin_samples then
+			max_bin_samples = bucket.count
 		end
 	end
-	return self;
+
+	print("");
+	print(prefix)
+	print(("_"):rep(graph_width*wscale).." "..max_bin_samples);
+	for row = graph_height, 1, -1 do
+		local row_chars = {};
+		local min_eighths, max_eighths = 8, 0;
+		for i = 1, #buckets do
+			local char_eighths = math.ceil(math.max(math.min((graph_height/(max_bin_samples/buckets[i].count))-(row-1), 1), 0)*8);
+			if char_eighths < min_eighths then
+				min_eighths = char_eighths;
+			end
+			if char_eighths > max_eighths then
+				max_eighths = char_eighths;
+			end
+			if char_eighths == 0 then
+				row_chars[i] = ("-"):rep(wscale);
+			else
+				local char = eighth_chars:sub(char_eighths*3+1, char_eighths*3+3);
+				row_chars[i] = char:rep(wscale);
+			end
+		end
+		print(table.concat(row_chars).."|- "..string.format("%.8g", math.ceil((max_bin_samples/graph_height)*(row-0.5))));
+	end
+
+	local legend_pat = string.format("%%%d.%dg", wscale-1, wscale-1)
+	local row = {}
+	for i = 1, #buckets do
+		local threshold = buckets[i].threshold
+		t_insert(row, legend_pat:format(threshold))
+	end
+	t_insert(row, " " .. metric_family.unit)
+	print(t_concat(row, "/"))
+
+	return true
 end
 
-function stats_methods:trim(lower, upper)
-	upper = upper or (100-lower);
-	local statistics = require "util.statistics";
-	for _, stat_info in ipairs(self) do
-		-- Strip outliers
-		local data = stat_info[4];
-		if data then
-			local new_data = {
-				min = statistics.get_percentile(data, lower);
-				max = statistics.get_percentile(data, upper);
-				samples = {};
-				sample_count = 0;
-				count = data.count;
-				units = data.units;
-			};
-			local sum = 0;
-			for _, v in ipairs(data.samples) do
-				if v > new_data.max then
-					break;
-				elseif v>=new_data.min then
-					table.insert(new_data.samples, v);
-					sum = sum + v;
-				end
-			end
-			new_data.sample_count = #new_data.samples;
-			stat_info[4] = new_data;
-			stat_info[3] = sum/new_data.sample_count;
-		end
-	end
-	return self;
+function stats_methods:render_single_fancy_histogram(print, prefix, metric_family, metric)
+	return self:render_single_fancy_histogram_ex(print, prefix, metric_family, metric, false)
 end
 
-function stats_methods:max(upper)
-	return self:bounds(nil, upper);
-end
-
-function stats_methods:min(lower)
-	return self:bounds(lower, nil);
-end
-
-function stats_methods:summary()
-	local statistics = require "util.statistics";
-	for _, stat_info in ipairs(self) do
-		local type, value, data = stat_info[2], stat_info[3], stat_info[4];
-		if data and data.samples then
-			table.insert(stat_info.output, string.format("Count: %d (%d captured)",
-				data.count,
-				data.sample_count
-			));
-			table.insert(stat_info.output, string.format("Min: %s  Mean: %s  Max: %s",
-				format_stat(type, data.units, data.min),
-				format_stat(type, data.units, value),
-				format_stat(type, data.units, data.max)
-			));
-			table.insert(stat_info.output, string.format("Q1: %s  Median: %s  Q3: %s",
-				format_stat(type, data.units, statistics.get_percentile(data, 25)),
-				format_stat(type, data.units, statistics.get_percentile(data, 50)),
-				format_stat(type, data.units, statistics.get_percentile(data, 75))
-			));
-		end
-	end
-	return self;
+function stats_methods:render_single_fancy_histogram_cf(print, prefix, metric_family, metric)
+	-- cf = cumulative frequency
+	return self:render_single_fancy_histogram_ex(print, prefix, metric_family, metric, true)
 end
 
 function stats_methods:cfgraph()
 	for _, stat_info in ipairs(self) do
-		local name, type, value, data = unpack(stat_info, 1, 4); -- luacheck: ignore 211
+		local family_name, metric_family = unpack(stat_info, 1, 2)
 		local function print(s)
 			table.insert(stat_info.output, s);
 		end
 
-		if data and data.sample_count and data.sample_count > 0 then
-			local raw_histogram = require "util.statistics".get_histogram(data);
-
-			local graph_width, graph_height = 50, 10;
-			local eighth_chars = "   ▁▂▃▄▅▆▇█";
-
-			local range = data.max - data.min;
-
-			if range > 0 then
-				local x_scaling = #raw_histogram/graph_width;
-				local histogram = {};
-				for i = 1, graph_width do
-					histogram[i] = math.max(raw_histogram[i*x_scaling-1] or 0, raw_histogram[i*x_scaling] or 0);
-				end
-
-				print("");
-				print(("_"):rep(52)..format_stat(type, data.units, data.max));
-				for row = graph_height, 1, -1 do
-					local row_chars = {};
-					local min_eighths, max_eighths = 8, 0;
-					for i = 1, #histogram do
-						local char_eighths = math.ceil(math.max(math.min((graph_height/(data.max/histogram[i]))-(row-1), 1), 0)*8);
-						if char_eighths < min_eighths then
-							min_eighths = char_eighths;
-						end
-						if char_eighths > max_eighths then
-							max_eighths = char_eighths;
-						end
-						if char_eighths == 0 then
-							row_chars[i] = "-";
-						else
-							local char = eighth_chars:sub(char_eighths*3+1, char_eighths*3+3);
-							row_chars[i] = char;
-						end
-					end
-					print(table.concat(row_chars).."|-"..format_stat(type, data.units, data.max/(graph_height/(row-0.5))));
-				end
-				print(("\\    "):rep(11));
-				local x_labels = {};
-				for i = 1, 11 do
-					local s = ("%-4s"):format((i-1)*10);
-					if #s > 4 then
-						s = s:sub(1, 3).."…";
-					end
-					x_labels[i] = s;
-				end
-				print(" "..table.concat(x_labels, " "));
-				local units = "%";
-				local margin = math.floor((graph_width-#units)/2);
-				print((" "):rep(margin)..units);
-			else
-				print("[range too small to graph]");
-			end
-			print("");
+		if not self:render_family(print, family_name, metric_family, self.render_single_fancy_histogram_cf) then
+			return self
 		end
 	end
 	return self;
@@ -1530,83 +1458,92 @@
 
 function stats_methods:histogram()
 	for _, stat_info in ipairs(self) do
-		local name, type, value, data = unpack(stat_info, 1, 4); -- luacheck: ignore 211
+		local family_name, metric_family = unpack(stat_info, 1, 2)
 		local function print(s)
 			table.insert(stat_info.output, s);
 		end
 
-		if not data then
-			print("[no data]");
-			return self;
-		elseif not data.sample_count then
-			print("[not a sampled metric type]");
-			return self;
+		if not self:render_family(print, family_name, metric_family, self.render_single_fancy_histogram) then
+			return self
 		end
-
-		local graph_width, graph_height = 50, 10;
-		local eighth_chars = "   ▁▂▃▄▅▆▇█";
-
-		local range = data.max - data.min;
-
-		if range > 0 then
-			local n_buckets = graph_width;
-
-			local histogram = {};
-			for i = 1, n_buckets do
-				histogram[i] = 0;
-			end
-			local max_bin_samples = 0;
-			for _, d in ipairs(data.samples) do
-				local bucket = math.floor(1+(n_buckets-1)/(range/(d-data.min)));
-				histogram[bucket] = histogram[bucket] + 1;
-				if histogram[bucket] > max_bin_samples then
-					max_bin_samples = histogram[bucket];
-				end
-			end
-
-			print("");
-			print(("_"):rep(52)..max_bin_samples);
-			for row = graph_height, 1, -1 do
-				local row_chars = {};
-				local min_eighths, max_eighths = 8, 0;
-				for i = 1, #histogram do
-					local char_eighths = math.ceil(math.max(math.min((graph_height/(max_bin_samples/histogram[i]))-(row-1), 1), 0)*8);
-					if char_eighths < min_eighths then
-						min_eighths = char_eighths;
-					end
-					if char_eighths > max_eighths then
-						max_eighths = char_eighths;
-					end
-					if char_eighths == 0 then
-						row_chars[i] = "-";
-					else
-						local char = eighth_chars:sub(char_eighths*3+1, char_eighths*3+3);
-						row_chars[i] = char;
-					end
-				end
-				print(table.concat(row_chars).."|-"..math.ceil((max_bin_samples/graph_height)*(row-0.5)));
-			end
-			print(("\\    "):rep(11));
-			local x_labels = {};
-			for i = 1, 11 do
-				local s = ("%-4s"):format(format_stat(type, data.units, data.min+range*i/11, data.min):match("^%S+"));
-				if #s > 4 then
-					s = s:sub(1, 3).."…";
-				end
-				x_labels[i] = s;
-			end
-			print(" "..table.concat(x_labels, " "));
-			local units = format_stat(type, data.units, data.min):match("%s+(.+)$") or data.units or "";
-			local margin = math.floor((graph_width-#units)/2);
-			print((" "):rep(margin)..units);
-		else
-			print("[range too small to graph]");
-		end
-		print("");
 	end
 	return self;
 end
 
+function stats_methods:render_single_counter(print, prefix, metric_family, metric)
+	local created_timestamp, current_value
+	for suffix, _, value in metric:iter_samples() do
+		if suffix == "_created" then
+			created_timestamp = value
+		elseif suffix == "_total" then
+			current_value = value
+		end
+	end
+	if current_value and created_timestamp then
+		local base_unit = short_units[metric_family.unit] or metric_family.unit
+		local unit = base_unit .. "/s"
+		local factor = 1
+		if base_unit == "s" then
+			-- be smart!
+			unit = "%"
+			factor = 100
+		elseif base_unit == "" then
+			unit = "events/s"
+		end
+		print(("%-50s %s"):format(prefix, format_number(factor * current_value / (self.now - created_timestamp), unit.." [avg]")));
+	end
+end
+
+function stats_methods:render_single_gauge(print, prefix, metric_family, metric)
+	local current_value
+	for _, _, value in metric:iter_samples() do
+		current_value = value
+	end
+	if current_value then
+		local unit = short_units[metric_family.unit] or metric_family.unit
+		print(("%-50s %s"):format(prefix, format_number(current_value, unit)));
+	end
+end
+
+function stats_methods:render_single_summary(print, prefix, metric_family, metric)
+	local sum, count
+	for suffix, _, value in metric:iter_samples() do
+		if suffix == "_sum" then
+			sum = value
+		elseif suffix == "_count" then
+			count = value
+		end
+	end
+	if sum and count then
+		local unit = short_units[metric_family.unit] or metric_family.unit
+		if count == 0 then
+			print(("%-50s %s"):format(prefix, "no obs."));
+		else
+			print(("%-50s %s"):format(prefix, format_number(sum / count, unit.."/event [avg]")));
+		end
+	end
+end
+
+function stats_methods:render_family(print, family_name, metric_family, render_func)
+	local labelkeys = metric_family.label_keys
+	if #labelkeys > 0 then
+		print(family_name)
+		for labelset, metric in metric_family:iter_metrics() do
+			local labels = {}
+			for i, k in ipairs(labelkeys) do
+				local v = labelset[i]
+				t_insert(labels, ("%s=%s"):format(k, v))
+			end
+			local prefix = "  "..t_concat(labels, " ")
+			render_func(self, print, prefix, metric_family, metric)
+		end
+	else
+		for _, metric in metric_family:iter_metrics() do
+			render_func(self, print, family_name, metric_family, metric)
+		end
+	end
+end
+
 local function stats_tostring(stats)
 	local print = stats.session.print;
 	for _, stat_info in ipairs(stats) do
@@ -1618,7 +1555,14 @@
 			end
 			print("");
 		else
-			print(("%-50s %s"):format(stat_info[1], format_stat(stat_info[2], (stat_info[4] or {}).units, stat_info[3])));
+			local metric_family = stat_info[2]
+			if metric_family.type_ == "counter" then
+				stats:render_family(print, stat_info[1], metric_family, stats.render_single_counter)
+			elseif metric_family.type_ == "gauge" or metric_family.type_ == "unknown" then
+				stats:render_family(print, stat_info[1], metric_family, stats.render_single_gauge)
+			elseif metric_family.type_ == "summary" or metric_family.type_ == "histogram" then
+				stats:render_family(print, stat_info[1], metric_family, stats.render_single_summary)
+			end
 		end
 	end
 	return #stats.." statistics displayed";
@@ -1626,23 +1570,29 @@
 
 local stats_mt = {__index = stats_methods, __tostring = stats_tostring }
 local function new_stats_context(self)
-	return setmetatable({ session = self.session, stats = true }, stats_mt);
+	-- TODO: instead of now(), it might be better to take the time of the last
+	-- interval, if the statistics backend is set to use periodic collection
+	-- Otherwise we get strange stuff like average cpu usage decreasing until
+	-- the next sample and so on.
+	return setmetatable({ session = self.session, stats = true, now = time.now() }, stats_mt);
 end
 
-function def_env.stats:show(filter)
-	-- luacheck: ignore 211/changed
-	local stats, changed, extra = require "core.statsmanager".get_stats();
-	local available, displayed = 0, 0;
+function def_env.stats:show(name_filter)
+	local statsman = require "core.statsmanager"
+	local collect = statsman.collect
+	if collect then
+		-- force collection if in manual mode
+		collect()
+	end
+	local metric_registry = statsman.get_metric_registry();
 	local displayed_stats = new_stats_context(self);
-	for name, value in iterators.sorted_pairs(stats) do
-		available = available + 1;
-		if not filter or name:match(filter) then
-			displayed = displayed + 1;
-			local type = name:match(":(%a+)$");
+	for family_name, metric_family in iterators.sorted_pairs(metric_registry:get_metric_families()) do
+		if not name_filter or family_name:match(name_filter) then
 			table.insert(displayed_stats, {
-				name, type, value, extra[name];
-				output = {};
-			});
+				family_name,
+				metric_family,
+				output = {}
+			})
 		end
 	end
 	return displayed_stats;
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/util/openmetrics.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -0,0 +1,308 @@
+--[[
+This module implements a subset of the OpenMetrics Internet Draft version 00.
+
+URL: https://tools.ietf.org/html/draft-richih-opsawg-openmetrics-00
+
+The following metric types are supported:
+
+- Counter
+- Gauge
+- Histogram
+- Summary
+
+It is used by util.statsd and util.statistics to provite the OpenMetrics API.
+
+To understand what this module is about, it is useful to familiarize oneself
+with the terms MetricFamily, Metric, LabelSet, Label and MetricPoint as
+defined in the I-D linked above.
+--]]
+-- metric constructor interface:
+-- metric_ctor(..., family_name, labels, extra)
+
+local time = require "util.time".now;
+local select = select;
+local array = require "util.array";
+local log = require "util.logger".init("util.openmetrics");
+local new_multitable = require "util.multitable".new;
+local iter_multitable = require "util.multitable".iter;
+
+-- BEGIN of Utility: "metric proxy"
+-- This allows to wrap a MetricFamily in a proxy which only provides the
+-- `with_labels` and `with_partial_label` methods. This allows to pre-set one
+-- or more labels on a metric family. This is used in particular via
+-- `with_partial_label` by the moduleapi in order to pre-set the `host` label
+-- on metrics created in non-global modules.
+local metric_proxy_mt = {}
+metric_proxy_mt.__index = metric_proxy_mt
+
+local function new_metric_proxy(metric_family, with_labels_proxy_fun)
+	return {
+		_family = metric_family,
+		with_labels = function(self, ...)
+			return with_labels_proxy_fun(self._family, ...)
+		end;
+		with_partial_label = function(self, label)
+			return new_metric_proxy(self._family, function(family, ...)
+				return family:with_labels(label, ...)
+			end)
+		end
+	}
+end
+
+-- END of Utility: "metric proxy"
+
+local function render_histogram_le(v)
+	if v == 1/0 then
+		-- I-D-00: 4.1.2.2.1:
+		--    Exposers MUST produce output for positive infinity as +Inf.
+		return "+Inf"
+	end
+
+	return string.format("%g", v)
+end
+
+-- BEGIN of generic MetricFamily implementation
+
+local metric_family_mt = {}
+metric_family_mt.__index = metric_family_mt
+
+local function histogram_metric_ctor(orig_ctor, buckets)
+	return function(family_name, labels, extra)
+		return orig_ctor(buckets, family_name, labels, extra)
+	end
+end
+
+local function new_metric_family(backend, type_, family_name, unit, description, label_keys, extra)
+	local metric_ctor = assert(backend[type_], "statistics backend does not support "..type_.." metrics families")
+	local labels = label_keys or {}
+	local user_labels = #labels
+	if type_ == "histogram" then
+		local buckets = extra and extra.buckets
+		if not buckets then
+			error("no buckets given for histogram metric")
+		end
+		buckets = array(buckets)
+		buckets:push(1/0)  -- must have +inf bucket
+
+		metric_ctor = histogram_metric_ctor(metric_ctor, buckets)
+	end
+
+	local data
+	if #labels == 0 then
+		data = metric_ctor(family_name, nil, extra)
+	else
+		data = new_multitable()
+	end
+
+	local mf = {
+		family_name = family_name,
+		data = data,
+		type_ = type_,
+		unit = unit,
+		description = description,
+		user_labels = user_labels,
+		label_keys = labels,
+		extra = extra,
+		_metric_ctor = metric_ctor,
+	}
+	setmetatable(mf, metric_family_mt);
+	return mf
+end
+
+function metric_family_mt:new_metric(labels)
+	return self._metric_ctor(self.family_name, labels, self.extra)
+end
+
+function metric_family_mt:clear()
+	for _, metric in self:iter_metrics() do
+		metric:reset()
+	end
+end
+
+function metric_family_mt:with_labels(...)
+	local count = select('#', ...)
+	if count ~= self.user_labels then
+		error("number of labels passed to with_labels does not match number of label keys")
+	end
+	if count == 0 then
+		return self.data
+	end
+	local metric = self.data:get(...)
+	if not metric then
+		local values = table.pack(...)
+		metric = self:new_metric(values)
+		values[values.n+1] = metric
+		self.data:set(table.unpack(values, 1, values.n+1))
+	end
+	return metric
+end
+
+function metric_family_mt:with_partial_label(label)
+	return new_metric_proxy(self, function (family, ...)
+		return family:with_labels(label, ...)
+	end)
+end
+
+function metric_family_mt:iter_metrics()
+	if #self.label_keys == 0 then
+		local done = false
+		return function()
+			if done then
+				return nil
+			end
+			done = true
+			return {}, self.data
+		end
+	end
+	local searchkeys = {};
+	local nlabels = #self.label_keys
+	for i=1,nlabels do
+		searchkeys[i] = nil;
+	end
+	local it, state = iter_multitable(self.data, table.unpack(searchkeys, 1, nlabels))
+	return function(_s)
+		local label_values = table.pack(it(_s))
+		if label_values.n == 0 then
+			return nil, nil
+		end
+		local metric = label_values[label_values.n]
+		label_values[label_values.n] = nil
+		label_values.n = label_values.n - 1
+		return label_values, metric
+	end, state
+end
+
+-- END of generic MetricFamily implementation
+
+-- BEGIN of MetricRegistry implementation
+
+
+-- Helper to test whether two metrics are "equal".
+local function equal_metric_family(mf1, mf2)
+	if mf1.type_ ~= mf2.type_ then
+		return false
+	end
+	if #mf1.label_keys ~= #mf2.label_keys then
+		return false
+	end
+	-- Ignoring unit here because in general it'll be part of the name anyway
+	-- So either the unit was moved into/out of the name (which is a valid)
+	-- thing to do on an upgrade or we would expect not to see any conflicts
+	-- anyway.
+	--[[
+	if mf1.unit ~= mf2.unit then
+		return false
+	end
+	]]
+	for i, key in ipairs(mf1.label_keys) do
+		if key ~= mf2.label_keys[i] then
+			return false
+		end
+	end
+	return true
+end
+
+-- If the unit is not empty, add it to the full name as per the I-D spec.
+local function compose_name(name, unit)
+	local full_name = name
+	if unit and unit ~= "" then
+		full_name = full_name .. "_" .. unit
+	end
+	-- TODO: prohibit certain suffixes used by metrics if where they may cause
+	-- conflicts
+	return full_name
+end
+
+local metric_registry_mt = {}
+metric_registry_mt.__index = metric_registry_mt
+
+local function new_metric_registry(backend)
+	local reg = {
+		families = {},
+		backend = backend,
+	}
+	setmetatable(reg, metric_registry_mt)
+	return reg
+end
+
+function metric_registry_mt:register_metric_family(name, metric_family)
+	local existing = self.families[name];
+	if existing then
+		if not equal_metric_family(metric_family, existing) then
+			-- We could either be strict about this, or replace the
+			-- existing metric family with the new one.
+			-- Being strict is nice to avoid programming errors /
+			-- conflicts, but causes issues when a new version of a module
+			-- is loaded.
+			--
+			-- We will thus assume that the new metric is the correct one;
+			-- That is probably OK because unless you're reaching down into
+			-- the util.openmetrics or core.statsmanager API, your metric
+			-- name is going to be scoped to `prosody_mod_$modulename`
+			-- anyway and the damage is thus controlled.
+			--
+			-- To make debugging such issues easier, we still log.
+			log("debug", "replacing incompatible existing metric family %s", name)
+			-- Below is the code to be strict.
+			--error("conflicting declarations for metric family "..name)
+		else
+			return existing
+		end
+	end
+	self.families[name] = metric_family
+	return metric_family
+end
+
+function metric_registry_mt:gauge(name, unit, description, labels, extra)
+	name = compose_name(name, unit)
+	local mf = new_metric_family(self.backend, "gauge", name, unit, description, labels, extra)
+	mf = self:register_metric_family(name, mf)
+	return mf
+end
+
+function metric_registry_mt:counter(name, unit, description, labels, extra)
+	name = compose_name(name, unit)
+	local mf = new_metric_family(self.backend, "counter", name, unit, description, labels, extra)
+	mf = self:register_metric_family(name, mf)
+	return mf
+end
+
+function metric_registry_mt:histogram(name, unit, description, labels, extra)
+	name = compose_name(name, unit)
+	local mf = new_metric_family(self.backend, "histogram", name, unit, description, labels, extra)
+	mf = self:register_metric_family(name, mf)
+	return mf
+end
+
+function metric_registry_mt:summary(name, unit, description, labels, extra)
+	name = compose_name(name, unit)
+	local mf = new_metric_family(self.backend, "summary", name, unit, description, labels, extra)
+	mf = self:register_metric_family(name, mf)
+	return mf
+end
+
+function metric_registry_mt:get_metric_families()
+	return self.families
+end
+
+-- END of MetricRegistry implementation
+
+-- BEGIN of general helpers for implementing high-level APIs on top of OpenMetrics
+
+local function timed(metric)
+	local t0 = time()
+	local submitter = assert(metric.sample or metric.set, "metric type cannot be used with timed()")
+	return function()
+		local t1 = time()
+		submitter(metric, t1-t0)
+	end
+end
+
+-- END of general helpers
+
+return {
+	new_metric_proxy = new_metric_proxy;
+	new_metric_registry = new_metric_registry;
+	render_histogram_le = render_histogram_le;
+	timed = timed;
+}
--- a/util/statistics.lua	Sun Apr 25 17:32:27 2021 +0200
+++ b/util/statistics.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -1,171 +1,191 @@
-local t_sort = table.sort
-local m_floor = math.floor;
 local time = require "util.time".now;
+local new_metric_registry = require "util.openmetrics".new_metric_registry;
+local render_histogram_le = require "util.openmetrics".render_histogram_le;
 
-local function nop_function() end
+-- BEGIN of Metric implementations
+
+-- Gauges
+local gauge_metric_mt = {}
+gauge_metric_mt.__index = gauge_metric_mt
+
+local function new_gauge_metric()
+	local metric = { value = 0 }
+	setmetatable(metric, gauge_metric_mt)
+	return metric
+end
+
+function gauge_metric_mt:set(value)
+	self.value = value
+end
+
+function gauge_metric_mt:add(delta)
+	self.value = self.value + delta
+end
+
+function gauge_metric_mt:reset()
+	self.value = 0
+end
 
-local function percentile(arr, length, pc)
-	local n = pc/100 * (length + 1);
-	local k, d = m_floor(n), n%1;
-	if k == 0 then
-		return arr[1] or 0;
-	elseif k >= length then
-		return arr[length];
-	end
-	return arr[k] + d*(arr[k+1] - arr[k]);
+function gauge_metric_mt:iter_samples()
+	local done = false
+	return function(_s)
+		if done then
+			return nil, true
+		end
+		done = true
+		return "", nil, _s.value
+	end, self
+end
+
+-- Counters
+local counter_metric_mt = {}
+counter_metric_mt.__index = counter_metric_mt
+
+local function new_counter_metric()
+	local metric = {
+		_created = time(),
+		value = 0,
+	}
+	setmetatable(metric, counter_metric_mt)
+	return metric
+end
+
+function counter_metric_mt:set(value)
+	self.value = value
+end
+
+function counter_metric_mt:add(value)
+	self.value = (self.value or 0) + value
+end
+
+function counter_metric_mt:iter_samples()
+	local step = 0
+	return function(_s)
+		step = step + 1
+		if step == 1 then
+			return "_created", nil, _s._created
+		elseif step == 2 then
+			return "_total", nil, _s.value
+		else
+			return nil, nil, true
+		end
+	end, self
+end
+
+function counter_metric_mt:reset()
+	self.value = 0
 end
 
-local function new_registry(config)
-	config = config or {};
-	local duration_sample_interval = config.duration_sample_interval or 5;
-	local duration_max_samples = config.duration_max_stored_samples or 5000;
+-- Histograms
+local histogram_metric_mt = {}
+histogram_metric_mt.__index = histogram_metric_mt
 
-	local function get_distribution_stats(events, n_actual_events, since, new_time, units)
-		local n_stored_events = #events;
-		t_sort(events);
-		local sum = 0;
-		for i = 1, n_stored_events do
-			sum = sum + events[i];
-		end
+local function new_histogram_metric(buckets)
+	local metric = {
+		_created = time(),
+		_sum = 0,
+		_count = 0,
+	}
+	-- the order of buckets matters unfortunately, so we cannot directly use
+	-- the thresholds as table keys
+	for i, threshold in ipairs(buckets) do
+		metric[i] = {
+			threshold = threshold,
+			threshold_s = render_histogram_le(threshold),
+			count = 0
+		}
+	end
+	setmetatable(metric, histogram_metric_mt)
+	return metric
+end
 
-		return {
-			samples = events;
-			sample_count = n_stored_events;
-			count = n_actual_events,
-			rate = n_actual_events/(new_time-since);
-			average = n_stored_events > 0 and sum/n_stored_events or 0,
-			min = events[1] or 0,
-			max = events[n_stored_events] or 0,
-			units = units,
-		};
+function histogram_metric_mt:sample(value)
+	-- According to the I-D, values must be part of all buckets
+	for i, bucket in pairs(self) do
+		if "number" == type(i) and bucket.threshold > value then
+			bucket.count = bucket.count + 1
+		end
 	end
+	self._sum = self._sum + value
+	self._count = self._count + 1
+end
 
+function histogram_metric_mt:iter_samples()
+	local key = nil
+	return function (_s)
+		local data
+		key, data = next(_s, key)
+		if key == "_created" or key == "_sum" or key == "_count" then
+			return key, nil, data
+		elseif key ~= nil then
+			return "_bucket", {["le"] = data.threshold_s}, data.count
+		else
+			return nil, nil, nil
+		end
+	end, self
+end
 
-	local registry = {};
-	local methods;
-	methods = {
-		amount = function (name, conf)
-			local v = conf and conf.initial or 0;
-			registry[name..":amount"] = function ()
-				return "amount", v, conf;
-			end
-			return function (new_v) v = new_v; end
-		end;
-		counter = function (name, conf)
-			local v = conf and conf.initial or 0;
-			registry[name..":amount"] = function ()
-				return "amount", v, conf;
-			end
-			return function (delta)
-				v = v + delta;
-			end;
-		end;
-		rate = function (name, conf)
-			local since, n, total = time(), 0, 0;
-			registry[name..":rate"] = function ()
-				total = total + n;
-				local t = time();
-				local stats = {
-					rate = n/(t-since);
-					count = n;
-					total = total;
-					units = conf and conf.units;
-					type = conf and conf.type;
-				};
-				since, n = t, 0;
-				return "rate", stats.rate, stats;
-			end;
-			return function ()
-				n = n + 1;
-			end;
-		end;
-		distribution = function (name, conf)
-			local units = conf and conf.units;
-			local type = conf and conf.type or "distribution";
-			local events, last_event = {}, 0;
-			local n_actual_events = 0;
-			local since = time();
+function histogram_metric_mt:reset()
+	self._created = time()
+	self._count = 0
+	self._sum = 0
+	for i, bucket in pairs(self) do
+		if "number" == type(i) then
+			bucket.count = 0
+		end
+	end
+end
+
+-- Summary
+local summary_metric_mt = {}
+summary_metric_mt.__index = summary_metric_mt
+
+local function new_summary_metric()
+	-- quantiles are not supported yet
+	local metric = {
+		_created = time(),
+		_sum = 0,
+		_count = 0,
+	}
+	setmetatable(metric, summary_metric_mt)
+	return metric
+end
 
-			registry[name..":"..type] = function ()
-				local new_time = time();
-				local stats = get_distribution_stats(events, n_actual_events, since, new_time, units);
-				events, last_event = {}, 0;
-				n_actual_events = 0;
-				since = new_time;
-				return type, stats.average, stats;
-			end;
+function summary_metric_mt:sample(value)
+	self._sum = self._sum + value
+	self._count = self._count + 1
+end
 
-			return function (value)
-				n_actual_events = n_actual_events + 1;
-				if n_actual_events%duration_sample_interval == 1 then
-					last_event = (last_event%duration_max_samples) + 1;
-					events[last_event] = value;
-				end
-			end;
-		end;
-		sizes = function (name, conf)
-			conf = conf or { units = "bytes", type = "size" }
-			return methods.distribution(name, conf);
-		end;
-		times = function (name, conf)
-			local units = conf and conf.units or "seconds";
-			local events, last_event = {}, 0;
-			local n_actual_events = 0;
-			local since = time();
+function summary_metric_mt:iter_samples()
+	local key = nil
+	return function (_s)
+		local data
+		key, data = next(_s, key)
+		return key, nil, data
+	end, self
+end
 
-			registry[name..":duration"] = function ()
-				local new_time = time();
-				local stats = get_distribution_stats(events, n_actual_events, since, new_time, units);
-				events, last_event = {}, 0;
-				n_actual_events = 0;
-				since = new_time;
-				return "duration", stats.average, stats;
-			end;
-
-			return function ()
-				n_actual_events = n_actual_events + 1;
-				if n_actual_events%duration_sample_interval ~= 1 then
-					return nop_function;
-				end
+function summary_metric_mt:reset()
+	self._created = time()
+	self._count = 0
+	self._sum = 0
+end
 
-				local start_time = time();
-				return function ()
-					local end_time = time();
-					local duration = end_time - start_time;
-					last_event = (last_event%duration_max_samples) + 1;
-					events[last_event] = duration;
-				end
-			end;
-		end;
+local pull_backend = {
+	gauge = new_gauge_metric,
+	counter = new_counter_metric,
+	histogram = new_histogram_metric,
+	summary = new_summary_metric,
+}
 
-		get_stats = function ()
-			return registry;
-		end;
-	};
-	return methods;
+-- END of Metric implementations
+
+local function new()
+	return {
+		metric_registry = new_metric_registry(pull_backend),
+	}
 end
 
 return {
-	new = new_registry;
-	get_histogram = function (duration, n_buckets)
-		n_buckets = n_buckets or 100;
-		local events, n_events = duration.samples, duration.sample_count;
-		if not (events and n_events) then
-			return nil, "not a valid distribution stat";
-		end
-		local histogram = {};
-
-		for i = 1, 100, 100/n_buckets do
-			histogram[i] = percentile(events, n_events, i);
-		end
-		return histogram;
-	end;
-
-	get_percentile = function (duration, pc)
-		local events, n_events = duration.samples, duration.sample_count;
-		if not (events and n_events) then
-			return nil, "not a valid distribution stat";
-		end
-		return percentile(events, n_events, pc);
-	end;
+	new = new;
 }
--- a/util/statsd.lua	Sun Apr 25 17:32:27 2021 +0200
+++ b/util/statsd.lua	Sun Apr 18 11:47:41 2021 +0200
@@ -1,6 +1,231 @@
 local socket = require "socket";
+local time = require "util.time".now;
+local array = require "util.array";
+local t_concat = table.concat;
+
+local new_metric_registry = require "util.openmetrics".new_metric_registry;
+local render_histogram_le = require "util.openmetrics".render_histogram_le;
+
+-- BEGIN of Metric implementations
+
+-- Gauges
+local gauge_metric_mt = {}
+gauge_metric_mt.__index = gauge_metric_mt
+
+local function new_gauge_metric(full_name, impl)
+	local metric = {
+		_full_name = full_name;
+		_impl = impl;
+		value = 0;
+	}
+	setmetatable(metric, gauge_metric_mt)
+	return metric
+end
+
+function gauge_metric_mt:set(value)
+	self.value = value
+	self._impl:push_gauge(self._full_name, value)
+end
+
+function gauge_metric_mt:add(delta)
+	self.value = self.value + delta
+	self._impl:push_gauge(self._full_name, self.value)
+end
+
+function gauge_metric_mt:reset()
+	self.value = 0
+	self._impl:push_gauge(self._full_name, 0)
+end
+
+function gauge_metric_mt.iter_samples()
+	-- statsd backend does not support iteration.
+	return function()
+		return nil
+	end
+end
+
+-- Counters
+local counter_metric_mt = {}
+counter_metric_mt.__index = counter_metric_mt
+
+local function new_counter_metric(full_name, impl)
+	local metric = {
+		_full_name = full_name,
+		_impl = impl,
+		value = 0,
+	}
+	setmetatable(metric, counter_metric_mt)
+	return metric
+end
+
+function counter_metric_mt:set(value)
+	local delta = value - self.value
+	self.value = value
+	self._impl:push_counter_delta(self._full_name, delta)
+end
+
+function counter_metric_mt:add(value)
+	self.value = (self.value or 0) + value
+	self._impl:push_counter_delta(self._full_name, value)
+end
+
+function counter_metric_mt.iter_samples()
+	-- statsd backend does not support iteration.
+	return function()
+		return nil
+	end
+end
+
+function counter_metric_mt:reset()
+	self.value = 0
+end
+
+-- Histograms
+local histogram_metric_mt = {}
+histogram_metric_mt.__index = histogram_metric_mt
+
+local function new_histogram_metric(buckets, full_name, impl)
+	-- NOTE: even though the more or less proprietrary dogstatsd has its own
+	-- histogram implementation, we push the individual buckets in this statsd
+	-- backend for both consistency and compatibility across statsd
+	-- implementations.
+	local metric = {
+		_sum_name = full_name..".sum",
+		_count_name = full_name..".count",
+		_impl = impl,
+		_created = time(),
+		_sum = 0,
+		_count = 0,
+	}
+	-- the order of buckets matters unfortunately, so we cannot directly use
+	-- the thresholds as table keys
+	for i, threshold in ipairs(buckets) do
+		local threshold_s = render_histogram_le(threshold)
+		metric[i] = {
+			threshold = threshold,
+			threshold_s = threshold_s,
+			count = 0,
+			_full_name = full_name..".bucket."..(threshold_s:gsub("%.", "_")),
+		}
+	end
+	setmetatable(metric, histogram_metric_mt)
+	return metric
+end
 
-local time = require "util.time".now
+function histogram_metric_mt:sample(value)
+	-- According to the I-D, values must be part of all buckets
+	for i, bucket in pairs(self) do
+		if "number" == type(i) and bucket.threshold > value then
+			bucket.count = bucket.count + 1
+			self._impl:push_counter_delta(bucket._full_name, 1)
+		end
+	end
+	self._sum = self._sum + value
+	self._count = self._count + 1
+	self._impl:push_gauge(self._sum_name, self._sum)
+	self._impl:push_counter_delta(self._count_name, 1)
+end
+
+function histogram_metric_mt.iter_samples()
+	-- statsd backend does not support iteration.
+	return function()
+		return nil
+	end
+end
+
+function histogram_metric_mt:reset()
+	self._created = time()
+	self._count = 0
+	self._sum = 0
+	for i, bucket in pairs(self) do
+		if "number" == type(i) then
+			bucket.count = 0
+		end
+	end
+	self._impl:push_gauge(self._sum_name, self._sum)
+end
+
+-- Summaries
+local summary_metric_mt = {}
+summary_metric_mt.__index = summary_metric_mt
+
+local function new_summary_metric(full_name, impl)
+	local metric = {
+		_sum_name = full_name..".sum",
+		_count_name = full_name..".count",
+		_impl = impl,
+	}
+	setmetatable(metric, summary_metric_mt)
+	return metric
+end
+
+function summary_metric_mt:sample(value)
+	self._impl:push_counter_delta(self._sum_name, value)
+	self._impl:push_counter_delta(self._count_name, 1)
+end
+
+function summary_metric_mt.iter_samples()
+	-- statsd backend does not support iteration.
+	return function()
+		return nil
+	end
+end
+
+function summary_metric_mt.reset()
+end
+
+-- BEGIN of statsd client implementation
+
+local statsd_mt = {}
+statsd_mt.__index = statsd_mt
+
+function statsd_mt:cork()
+	self.corked = true
+	self.cork_buffer = self.cork_buffer or {}
+end
+
+function statsd_mt:uncork()
+	self.corked = false
+	self:_flush_cork_buffer()
+end
+
+function statsd_mt:_flush_cork_buffer()
+	local buffer = self.cork_buffer
+	for metric_name, value in pairs(buffer) do
+		self:_send_gauge(metric_name, value)
+		buffer[metric_name] = nil
+	end
+end
+
+function statsd_mt:push_gauge(metric_name, value)
+	if self.corked then
+		self.cork_buffer[metric_name] = value
+	else
+		self:_send_gauge(metric_name, value)
+	end
+end
+
+function statsd_mt:_send_gauge(metric_name, value)
+	self:_send(self.prefix..metric_name..":"..tostring(value).."|g")
+end
+
+function statsd_mt:push_counter_delta(metric_name, delta)
+	self:_send(self.prefix..metric_name..":"..tostring(delta).."|c")
+end
+
+function statsd_mt:_send(s)
+	return self.sock:send(s)
+end
+
+-- END of statsd client implementation
+
+local function build_metric_name(family_name, labels)
+	local parts = array { family_name }
+	if labels then
+		parts:append(labels)
+	end
+	return t_concat(parts, "/"):gsub("%.", "_"):gsub("/", ".")
+end
 
 local function new(config)
 	if not config or not config.statsd_server then
@@ -12,71 +237,31 @@
 
 	local prefix = (config.prefix or "prosody")..".";
 
-	local function send_metric(s)
-		return sock:send(prefix..s);
-	end
-
-	local function send_gauge(name, amount, relative)
-		local s_amount = tostring(amount);
-		if relative and amount > 0 then
-			s_amount = "+"..s_amount;
-		end
-		return send_metric(name..":"..s_amount.."|g");
-	end
-
-	local function send_counter(name, amount)
-		return send_metric(name..":"..tostring(amount).."|c");
-	end
-
-	local function send_duration(name, duration)
-		return send_metric(name..":"..tostring(duration).."|ms");
-	end
-
-	local function send_histogram_sample(name, sample)
-		return send_metric(name..":"..tostring(sample).."|h");
-	end
+	local impl = {
+		metric_registry = nil;
+		sock = sock;
+		prefix = prefix;
+	};
+	setmetatable(impl, statsd_mt)
 
-	local methods;
-	methods = {
-		amount = function (name, conf)
-			if conf and conf.initial then
-				send_gauge(name, conf.initial);
-			end
-			return function (new_v) send_gauge(name, new_v); end
-		end;
-		counter = function (name, conf) --luacheck: ignore 212/conf
-			return function (delta)
-				send_gauge(name, delta, true);
-			end;
-		end;
-		rate = function (name)
-			return function ()
-				send_counter(name, 1);
-			end;
+	local backend = {
+		gauge = function(family_name, labels)
+			return new_gauge_metric(build_metric_name(family_name, labels), impl)
 		end;
-		distribution = function (name, conf) --luacheck: ignore 212/conf
-			return function (value)
-				send_histogram_sample(name, value);
-			end;
+		counter = function(family_name, labels)
+			return new_counter_metric(build_metric_name(family_name, labels), impl)
 		end;
-		sizes = function (name)
-			name = name.."_size";
-			return function (value)
-				send_histogram_sample(name, value);
-			end;
+		histogram = function(buckets, family_name, labels)
+			return new_histogram_metric(buckets, build_metric_name(family_name, labels), impl)
 		end;
-		times = function (name)
-			return function ()
-				local start_time = time();
-				return function ()
-					local end_time = time();
-					local duration = end_time - start_time;
-					send_duration(name, duration*1000);
-				end
-			end;
+		summary = function(family_name, labels, extra)
+			return new_summary_metric(build_metric_name(family_name, labels), impl, extra)
 		end;
 	};
-	return methods;
+
+	impl.metric_registry = new_metric_registry(backend);
+
+	return impl;
 end
 
 return {