cassandra-pr mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <>
Subject [GitHub] jolynch commented on a change in pull request #283: CASSANDRA-14459: DynamicEndpointSnitch should never prefer latent replicas
Date Thu, 24 Jan 2019 22:17:27 GMT
jolynch commented on a change in pull request #283: CASSANDRA-14459: DynamicEndpointSnitch
should never prefer latent replicas

 File path: src/java/org/apache/cassandra/locator/
 @@ -21,133 +21,505 @@
 import java.util.*;
+import java.util.concurrent.CancellationException;
 import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.ExecutionException;
 import java.util.concurrent.ScheduledFuture;
 import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
+import java.util.concurrent.atomic.AtomicLong;
-import com.codahale.metrics.ExponentiallyDecayingReservoir;
-import com.codahale.metrics.Snapshot;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import org.apache.cassandra.concurrent.DebuggableScheduledThreadPoolExecutor;
 import org.apache.cassandra.concurrent.ScheduledExecutors;
 import org.apache.cassandra.config.DatabaseDescriptor;
+import org.apache.cassandra.exceptions.RequestFailureReason;
 import org.apache.cassandra.gms.ApplicationState;
 import org.apache.cassandra.gms.EndpointState;
 import org.apache.cassandra.gms.Gossiper;
 import org.apache.cassandra.gms.VersionedValue;
+import org.apache.cassandra.locator.dynamicsnitch.DynamicEndpointSnitchHistogram;
 import org.apache.cassandra.service.StorageService;
 import org.apache.cassandra.utils.FBUtilities;
 import org.apache.cassandra.utils.MBeanWrapper;
+import static;
+import static;
+import static;
  * A dynamic snitch that sorts endpoints by latency with an adapted phi failure detector
+ * Note that the subclasses (e.g. {@link DynamicEndpointSnitchHistogram}) are responsible
for actually measuring
+ * latency and providing an ISnitchMeasurement implementation back to this class.
-public class DynamicEndpointSnitch extends AbstractEndpointSnitch implements ILatencySubscriber,
+public abstract class DynamicEndpointSnitch extends AbstractEndpointSnitch implements ILatencySubscriber,
-    private static final boolean USE_SEVERITY = !Boolean.getBoolean("cassandra.ignore_dynamic_snitch_severity");
+    private static final Logger logger = LoggerFactory.getLogger(DynamicEndpointSnitch.class);
+    // Latency measurement and ranking. The samples contain latency measurements and the
scores are used for ranking
+    protected boolean registered = false;
+    protected static final boolean USE_SEVERITY = !Boolean.getBoolean("cassandra.ignore_dynamic_snitch_severity");
+    protected volatile Map<InetAddressAndPort, Double> scores = new HashMap<>();
-    private static final double ALPHA = 0.75; // set to 0.75 to make EDS more biased to towards
the newer values
-    private static final int WINDOW_SIZE = 100;
+    protected final Map<InetAddressAndPort, AnnotatedMeasurement> samples = new ConcurrentHashMap<>();
-    private volatile int dynamicUpdateInterval = DatabaseDescriptor.getDynamicUpdateInterval();
-    private volatile int dynamicResetInterval = DatabaseDescriptor.getDynamicResetInterval();
-    private volatile double dynamicBadnessThreshold = DatabaseDescriptor.getDynamicBadnessThreshold();
+    // Latency probe functionality for actively probing endpoints that we haven't measured
recently but are ranking
+    public static final long MAX_PROBE_INTERVAL_MS = Long.getLong("cassandra.dynamic_snitch_max_probe_interval_ms",
60 * 10 * 1000L);
+    public static final long MIN_PROBE_INTERVAL_MS = Long.getLong("cassandra.dynamic_snitch_min_probe_interval_ms",
60 * 1000L) ;
+    // The probe rate is set later when configuration is read in applyConfigChanges
+    protected static final RateLimiter probeRateLimiter = RateLimiter.create(1);
+    protected static final DebuggableScheduledThreadPoolExecutor latencyProbeExecutor = new
+    private long lastUpdateSamplesNanos = System.nanoTime();
+    // User configuration of the snitch tunables
+    protected volatile int dynamicUpdateInterval = -1;
+    protected volatile int dynamicSampleUpdateInterval = -1;
+    protected volatile double dynamicBadnessThreshold = 0;
     // the score for a merged set of endpoints must be this much worse than the score for
separate endpoints to
     // warrant not merging two ranges into a single range
     private static final double RANGE_MERGING_PREFERENCE = 1.5;
     private String mbeanName;
-    private boolean registered = false;
-    private volatile HashMap<InetAddressAndPort, Double> scores = new HashMap<>();
-    private final ConcurrentHashMap<InetAddressAndPort, ExponentiallyDecayingReservoir>
samples = new ConcurrentHashMap<>();
+    private boolean mbeanRegistered = false;
     public final IEndpointSnitch subsnitch;
-    private volatile ScheduledFuture<?> updateSchedular;
-    private volatile ScheduledFuture<?> resetSchedular;
-    private final Runnable update;
-    private final Runnable reset;
+    private volatile ScheduledFuture<?> updateScoresScheduler = null;
+    private volatile ScheduledFuture<?> updateSamplesScheduler = null;
     public DynamicEndpointSnitch(IEndpointSnitch snitch)
         this(snitch, null);
-    public DynamicEndpointSnitch(IEndpointSnitch snitch, String instance)
+    protected DynamicEndpointSnitch(IEndpointSnitch snitch, String instance)
         mbeanName = "org.apache.cassandra.db:type=DynamicEndpointSnitch";
         if (instance != null)
             mbeanName += ",instance=" + instance;
         subsnitch = snitch;
-        update = new Runnable()
+        if (DatabaseDescriptor.isDaemonInitialized())
-            public void run()
+            open();
+        }
+    }
+    /**
+     * Update configuration of the background tasks and restart the various scheduler tasks
+     * if the configured rates for these tasks have changed.
+     */
+    public void applyConfigChanges(int newDynamicUpdateInternal, int newDynamicSampleUpdateInterval,
double newDynamicBadnessThreshold)
+    {
+        if (DatabaseDescriptor.isDaemonInitialized())
+        {
+            if (dynamicUpdateInterval != newDynamicUpdateInternal || updateScoresScheduler
== null)
-                updateScores();
+                cancelAndWait(updateScoresScheduler, Math.max(1, dynamicUpdateInterval),
 Review comment:
   Interesting, I usually opt to either not block or block with a timeout, but I see your
point. In this case none of the futures that are being called should block, and it can't block
during the initial startup (at constructor time the futures are `null` so they just immediately
succeed at cancelling) so I think I'll just block forever although it pains me to do so heh.
   I tried just weaving the TimeoutException back up the stack so we could have our cake and
eat it too, but there are lots of places that handling it just becomes difficult (e.g. in
the backwards compatible mbean interface I'd have to re-throw it as a runtime exception or
something which is confusing and not clear that your DES is now in a bad state).

This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:

With regards,
Apache Git Services

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message