3 files changed, 34 insertions(+), 10 deletions(-)

M config/certmon.sample.config
M src/cert_scanner.erl
M src/metrics_handler.erl

@@ 1,5 1,6 @@
 {port, 9101}.
 {scan_interval, 30000}.
+{max_error_count, 10}.
 {targets, [
            {"neo.sinyax.net", 443}
           ]

@@ 18,7 18,11 @@
           hostname,
           port,
           timer,
-          validity
+          validity,
+          % count of errors since last success
+          errorCount,
+          % total count of errors
+          totalErrors
 }).
 
 %% API.


@@ 38,11 42,16 @@ init([{Hostname, Port}]) ->
           hostname = Hostname,
           port = Port,
           timer = Timer,
-          validity = unknown
+          validity = unknown,
+          errorCount = 0,
+          totalErrors = 0
          }}.
 
 handle_call({get_status}, _From, State) ->
-  {reply, State#state.validity, State};
+  {reply, [
+           State#state.validity,
+           {errors, State#state.totalErrors}
+          ], State};
 handle_call(_Request, _From, State) ->
   {reply, ignored, State}.
 


@@ 83,10 92,22 @@ do_scan(State) ->
       RemainingDays = (NotAfterSeconds - NowSeconds) / (24 * 3600),
       logger:debug("peer cert ~s valid until: ~p (unix epoch: ~p, remaining days: ~p)~n",
                    [Hostname, NotAfter, NotAfterEpoch, RemainingDays]),
-      {noreply, State#state{validity = {valid, NotAfterEpoch}}};
+      {noreply, State#state{validity = {valid, NotAfterEpoch}, errorCount = 0}};
     {error, Reason} ->
       logger:warning("failed to connect to ~p: ~p", [Hostname, Reason]),
-      {noreply, State#state{validity = {error, Reason}}}
+      {ok, MaxErrorCount} = application:get_env(max_error_count),
+
+      % if max error count is not reached, we keep the old validity
+      ReportedValidity = if
+                           State#state.errorCount < MaxErrorCount ->
+                             State#state.validity;
+                           true -> {error, Reason}
+                         end,
+      {noreply, State#state{
+                  validity = ReportedValidity,
+                  errorCount = State#state.errorCount + 1,
+                  totalErrors = State#state.totalErrors + 1
+      }}
   end.
 
 %% utilities

@@ 6,16 6,18 @@
 format_metric({Hostname, Port}) ->
   Target = {Hostname, Port},
   {ok, ScannerPid} = certmon_sup:get_scanner_pid(Target),
-  Valid = try gen_server:call(ScannerPid, {get_status}) of
-            {valid, Epoch} -> Epoch;
-            _ -> 0
+  Response = try gen_server:call(ScannerPid, {get_status}) of
+            R -> R
           catch
             exit:Reason ->
               logger:error("failed to get status for ~p: Reason: ~p", [{Hostname, Port}, Reason]),
               0
           end,
-  [<<"tls_cert_expiry{hostname=\"">>, Hostname, <<"\", port=\"">>,
-   io_lib:format("~p", [Port]), <<"\"} ">>, io_lib:format("~p\n", [Valid])
+  Valid = proplists:get_value(valid, Response, 0),
+  Errors = proplists:get_value(errors, Response, 0),
+  [
+   io_lib:format("tls_cert_expiry{hostname=\"~p\", port=\"~p\"} ~p\n", [Hostname, Port, Valid]),
+   io_lib:format("tls_cert_errors{hostname=\"~p\", port=\"~p\"} ~p\n", [Hostname, Port, Errors])
   ].
 
 init(Req0, State) ->