[6.5][ML] Fixes potentials for SIGSEGV in multivariate probability calculation (#172)

Backport of #170.

Author: tveasey

docs/CHANGELOG.asciidoc

@@ -48,6 +48,7 @@ Persist and restore was missing some of the trend model state ({pull}#99[#99])
 Stop zero variance data generating a log error in the forecast confidence interval calculation ({pull}#107[#107])
 Fix corner case failing to calculate lgamma values and the correspoinding log errors ({pull}#126[#126])
 Influence count per bucket for metric population analyses was wrong and lead to wrong influencer scoring ({pull}#150[#150])
+Fix a possible SIGSEGV for jobs with multivariate by fields enabled which would lead to the job failing ({pull}#170[#170])
 
 === Regressions
 

include/maths/CTimeSeriesModel.h

@@ -247,7 +247,23 @@ class MATHS_EXPORT CUnivariateTimeSeriesModel : public CModel {
 
     //! Reinitialize state after detecting a new component of the trend
     //! decomposition.
-    void reinitializeStateGivenNewComponent(void);
+    void reinitializeStateGivenNewComponent();
+
+    //! Compute the probability for uncorrelated series.
+    bool uncorrelatedProbability(const CModelProbabilityParams& params,
+                                 const TTime2Vec1Vec& time,
+                                 const TDouble2Vec1Vec& value,
+                                 double& probability,
+                                 TTail2Vec& tail) const;
+
+    //! Compute the probability for correlated series.
+    bool correlatedProbability(const CModelProbabilityParams& params,
+                               const TTime2Vec1Vec& time,
+                               const TDouble2Vec1Vec& value,
+                               double& probability,
+                               TTail2Vec& tail,
+                               bool& conditional,
+                               TSize1Vec& mostAnomalousCorrelate) const;
 
     //! Get the models for the correlations and the models of the correlated
     //! time series.
@@ -458,9 +474,12 @@ class MATHS_EXPORT CTimeSeriesCorrelations {
     //! Add the time series identified by \p id.
     void addTimeSeries(std::size_t id, const CUnivariateTimeSeriesModel& model);
 
-    //! Remove the correlates of \p id.
+    //! Remove the time series identified by \p id.
     void removeTimeSeries(std::size_t id);
 
+    //! Clear all correlation information for time series identified \p id.
+    void clearCorrelationModels(std::size_t id);
+
     //! Add a sample for the time series identified by \p id.
     void addSamples(std::size_t id,
                     const CModelAddSamplesParams& params,

lib/maths/CMultimodalPrior.cc

@@ -139,7 +139,7 @@ bool CMultimodalPrior::acceptRestoreTraverser(const SDistributionRestoreParams&
                                this->numberSamples(numberSamples))
     } while (traverser.next());
 
-    if (m_Clusterer) {
+    if (m_Clusterer != nullptr) {
         // Register the split and merge callbacks.
         m_Clusterer->splitFunc(CModeSplitCallback(*this));
         m_Clusterer->mergeFunc(CModeMergeCallback(*this));
@@ -150,11 +150,13 @@ bool CMultimodalPrior::acceptRestoreTraverser(const SDistributionRestoreParams&
 
 CMultimodalPrior::CMultimodalPrior(const CMultimodalPrior& other)
     : CPrior(other.dataType(), other.decayRate()),
-      m_Clusterer(other.m_Clusterer->clone()),
-      m_SeedPrior(other.m_SeedPrior->clone()) {
+      m_Clusterer(other.m_Clusterer != nullptr ? other.m_Clusterer->clone() : nullptr),
+      m_SeedPrior(other.m_SeedPrior != nullptr ? other.m_SeedPrior->clone() : nullptr) {
     // Register the split and merge callbacks.
-    m_Clusterer->splitFunc(CModeSplitCallback(*this));
-    m_Clusterer->mergeFunc(CModeMergeCallback(*this));
+    if (m_Clusterer != nullptr) {
+        m_Clusterer->splitFunc(CModeSplitCallback(*this));
+        m_Clusterer->mergeFunc(CModeMergeCallback(*this));
+    }
 
     // Clone all the modes up front so we can implement strong exception safety.
     TModeVec modes;
@@ -182,10 +184,14 @@ void CMultimodalPrior::swap(CMultimodalPrior& other) {
     // The call backs for split and merge should point to the
     // appropriate priors (we don't swap the "this" pointers
     // after all). So we need to refresh them after swapping.
-    m_Clusterer->splitFunc(CModeSplitCallback(*this));
-    m_Clusterer->mergeFunc(CModeMergeCallback(*this));
-    other.m_Clusterer->splitFunc(CModeSplitCallback(other));
-    other.m_Clusterer->mergeFunc(CModeMergeCallback(other));
+    if (m_Clusterer != nullptr) {
+        m_Clusterer->splitFunc(CModeSplitCallback(*this));
+        m_Clusterer->mergeFunc(CModeMergeCallback(*this));
+    }
+    if (other.m_Clusterer != nullptr) {
+        other.m_Clusterer->splitFunc(CModeSplitCallback(other));
+        other.m_Clusterer->mergeFunc(CModeMergeCallback(other));
+    }
 
     std::swap(m_SeedPrior, other.m_SeedPrior);
     m_Modes.swap(other.m_Modes);

lib/maths/CMultivariateOneOfNPrior.cc

@@ -23,6 +23,7 @@
 #include <maths/Constants.h>
 
 #include <boost/bind.hpp>
+#include <boost/make_unique.hpp>
 #include <boost/numeric/conversion/bounds.hpp>
 #include <boost/ref.hpp>
 
@@ -438,7 +439,7 @@ CMultivariateOneOfNPrior::univariate(const TSize10Vec& marginalize,
         models[i].first *= std::exp(weights[i] - maxWeight[0]) / Z;
     }
 
-    return {TUnivariatePriorPtr(new COneOfNPrior(models, this->dataType(), this->decayRate())),
+    return {boost::make_unique<COneOfNPrior>(models, this->dataType(), this->decayRate()),
             maxWeight.count() > 0 ? maxWeight[0] : 0.0};
 }
 
@@ -471,8 +472,8 @@ CMultivariateOneOfNPrior::bivariate(const TSize10Vec& marginalize,
         models[i].first *= std::exp(weights[i] - maxWeight[0]) / Z;
     }
 
-    return {TPriorPtr(new CMultivariateOneOfNPrior(2, models, this->dataType(),
-                                                   this->decayRate())),
+    return {boost::make_unique<CMultivariateOneOfNPrior>(2, models, this->dataType(),
+                                                         this->decayRate()),
             maxWeight.count() > 0 ? maxWeight[0] : 0.0};
 }