Introduce Checkable#scheduler_shuffle_cap

Author: Al2Klimov

doc/09-object-types.md

@@ -352,6 +352,7 @@ Configuration Attributes:
   check\_timeout            | Duration              | **Optional.** Check command timeout in seconds. Overrides the CheckCommand's `timeout` attribute.
   check\_interval           | Duration              | **Optional.** The check interval (in seconds). This interval is used for checks when the host is in a `HARD` state. Defaults to `5m`.
   retry\_interval           | Duration              | **Optional.** The retry interval (in seconds). This interval is used for checks when the host is in a `SOFT` state. Defaults to `1m`. Note: This does not affect the scheduling [after a passive check result](08-advanced-topics.md#check-result-freshness).
+  scheduler\_shuffle\_cap   | Number                | **Optional.** Number of percent by up to which Icinga is allowed to override the check interval arbitrarily and in any direction to reduce load spikes. Defaults to 0.
   enable\_notifications     | Boolean               | **Optional.** Whether notifications are enabled. Defaults to true.
   enable\_active\_checks    | Boolean               | **Optional.** Whether active checks are enabled. Defaults to true.
   enable\_passive\_checks   | Boolean               | **Optional.** Whether passive checks are enabled. Defaults to true.
@@ -719,6 +720,7 @@ Configuration Attributes:
   check\_timeout            | Duration              | **Optional.** Check command timeout in seconds. Overrides the CheckCommand's `timeout` attribute.
   check\_interval           | Duration              | **Optional.** The check interval (in seconds). This interval is used for checks when the service is in a `HARD` state. Defaults to `5m`.
   retry\_interval           | Duration              | **Optional.** The retry interval (in seconds). This interval is used for checks when the service is in a `SOFT` state. Defaults to `1m`. Note: This does not affect the scheduling [after a passive check result](08-advanced-topics.md#check-result-freshness).
+  scheduler\_shuffle\_cap   | Number                | **Optional.** Number of percent by up to which Icinga is allowed to override the check interval arbitrarily and in any direction to reduce load spikes. Defaults to 0.
   enable\_notifications     | Boolean               | **Optional.** Whether notifications are enabled. Defaults to `true`.
   enable\_active\_checks    | Boolean               | **Optional.** Whether active checks are enabled. Defaults to `true`.
   enable\_passive\_checks   | Boolean               | **Optional.** Whether passive checks are enabled. Defaults to `true`.

lib/icinga/checkable-check.cpp

@@ -14,6 +14,7 @@
 #include "base/convert.hpp"
 #include "base/utility.hpp"
 #include "base/context.hpp"
+#include <cstdlib>
 
 using namespace icinga;
 
@@ -67,7 +68,7 @@ void Checkable::UpdateNextCheck(const MessageOrigin::Ptr& origin)
 	if (adj != 0.0)
 		adj = std::min(0.5 + fmod(GetSchedulingOffset(), interval * 5) / 100.0, adj);
 
-	double nextCheck = now - adj + interval;
+	double nextCheck = now - adj + interval * GetIntervalShuffleFactor();
 	double lastCheck = GetLastCheck();
 
 	Log(LogDebug, "Checkable")
@@ -372,7 +373,7 @@ Checkable::ProcessingResult Checkable::ProcessCheckResult(const CheckResult::Ptr
 		if (ttl > 0)
 			offset = ttl;
 		else
-			offset = GetCheckInterval();
+			offset = GetCheckInterval() * GetIntervalShuffleFactor();
 
 		SetNextCheck(Utility::GetTime() + offset, false, origin);
 	}
@@ -412,7 +413,7 @@ Checkable::ProcessingResult Checkable::ProcessCheckResult(const CheckResult::Ptr
 			if (!parent->GetEnableActiveChecks())
 				continue;
 
-			if (parent->GetNextCheck() >= now + parent->GetRetryInterval()) {
+			if (parent->GetNextCheck() >= now + parent->GetRetryInterval() * parent->GetIntervalShuffleFactor()) {
 				ObjectLock olock(parent);
 				parent->SetNextCheck(now);
 			}
@@ -707,3 +708,20 @@ void Checkable::AquirePendingCheckSlot(int maxPendingChecks)
 
 	m_PendingChecks++;
 }
+
+/**
+ * Returns a random factor derived from scheduler_shuffle_cap to multiply the check interval with.
+ *
+ * E.g. if scheduler_shuffle_cap is 20 (%), this function returns [0.8, 1.2].
+ */
+double Checkable::GetIntervalShuffleFactor()
+{
+	if (!GetEnableActiveChecks()) {
+		// scheduler_shuffle_cap doesn't influence external checkers.
+		return 1;
+	}
+
+	return (GetSchedulerShuffleCap() / 100) // scheduler_shuffle_cap as non-%, i.e. 10 => 0.1
+		* (rand() / (double)RAND_MAX * 2 - 1) // random number [-1, 1]
+		+ 1;
+}

lib/icinga/checkable.cpp

@@ -9,6 +9,7 @@
 #include "base/exception.hpp"
 #include "base/timer.hpp"
 #include <boost/thread/once.hpp>
+#include <cmath>
 
 using namespace icinga;
 
@@ -91,9 +92,9 @@ void Checkable::Start(bool runtimeCreated)
 	}
 
 	if (GetNextCheck() < now + 60) {
-		double delta = std::min(GetCheckInterval(), 60.0);
+		double delta = std::min(GetCheckInterval() * GetIntervalShuffleFactor(), 60.0);
 		delta *= (double)std::rand() / RAND_MAX;
-		SetNextCheck(now + delta);
+		SetNextCheck(now + delta + GetCheckInterval() * fabs(GetIntervalShuffleFactor() - 1));
 	}
 
 	ObjectImpl<Checkable>::Start(runtimeCreated);

lib/icinga/checkable.hpp

@@ -198,6 +198,7 @@ class Checkable : public ObjectImpl<Checkable>
 	bool NotificationReasonApplies(NotificationType type);
 	bool NotificationReasonSuppressed(NotificationType type);
 	bool IsLikelyToBeCheckedSoon();
+	double GetIntervalShuffleFactor();
 
 	void FireSuppressedNotifications();
 

lib/icinga/checkable.ti

@@ -47,6 +47,9 @@ abstract class Checkable : CustomVarObject
 	[config] double retry_interval {
 		default {{{ return 60; }}}
 	};
+	[config] double scheduler_shuffle_cap {
+		default {{{ return 0.0; }}}
+	};
 	[config, navigation] name(EventCommand) event_command (EventCommandRaw) {
 		navigate {{{
 			return EventCommand::GetByName(GetEventCommandRaw());