This patch adds the implementation for begin and finish local sampling
of a column family.
There is a difference in the implementation of Cassandra API and Scylla.
In Cassandra and the JMX an external source start and stop the sampling.
In Scylla, a single API call start the sampling and return with the
result. In Scylla the API call always return sampling of the read and of
the writes.
To bridge the difference, the begin sampling command will use a Future
when calling the API. The finish method will wait for the future to end.
Because of the different implementation, it is possible that two
consecutive calls will be made to start sampling one for the read and
one for the write, similarly, two calls will be made to finish for read
and write.
The implementation would ignore the second call to start and will
store the result, so the second call to finish will be served from the
stored result.
Note, that the use of future is only for safety, the way we expect it to
work, the caller to the begin sampling will sleep anyhow while waiting
for the result.
To avoid breaking the MBean compatibility we piggyback the duration on
top of the sampler string.
If no duration is given, a default duration will be taken, this is also
just as a precaution, we will modify the nodetool implementation to
pass that information.
There is a known issue with cardinality, that will need to be addressed.
Also we return a value in the raw column to match what Cassandra JMX
returns, but it's a duplication of the partition key.
See scylladb/scylla#2811
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Message-Id: <20190128143505.5241-1-amnon@scylladb.com>
Fixes#57
The usage of TableMetricsObjectName-yada-yada relies on translating the
"fake" objectname to a canonical one on remote
publication/serialization. However, the implementation of
ObjectName.getInstance has changed in JDK (micro) updates so it no
longer applies overridable methods -> wrong name published.
Fix by doing explicit ObjectName instansiation.
Message-Id: <20181023132005.23099-1-calle@scylladb.com>
Almost 100% null implementations, which is ok for most purposes
currently used by scylla. Some of these new calls (like dropped
mutations etc) should perhaps however be implemented.
Tested with the nodetool dtests. So sparsely.
Needed when/if scylla-tools-java is upgraded to origin 3.11,
otherwise noodtool breaks.
Message-Id: <20180730113741.14952-1-calle@scylladb.com>
This is a new extention of ObjectName that uses less memory.
TableMetricNameFactory and AllTableMetricNameFactory can
create it instead of regular ObjectName to save memory.
It is possible to save memory because each name created by
TableMetricNameFactory (or AllTableMetricNameFactory) shares
most of its data with other names created by the same factory
and there's no need to create multiple copies.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
JmxMBeanServer is a concrete implementation of a MBeanServer.
We want to use it directly because we need to bypass calls to
JmxMBeanServer.registerMBean and JmxMBeanServer.unregisterMBean.
They take ObjectName as parameter, copy it using
ObjectName.getInstance(ObjectName) and pass it to registerMBean
and unregisterMBean of JmxMBeanServer.getMBeanServerInterceptor().
We want to avoid this copy and pass the ObjectName argument directly
to JmxMBeanServer.getMBeanServerInterceptor().
To do that this patch:
1. changes all MBeanServer variables to JmxMBeanServer
2. creates JmxMBeanServer in APIBuilder making sure accessing
interceptors is allowed
3. makes sure that JmxMBeanServer.getMBeanServerInterceptor().registerMBean
is called wherever JmxMBeanServer.registerMBean was called
4. makes sure that JmxMBeanServer.getMBeanServerInterceptor().unregisterMBean
is called whenever JmxMBeanServer.unregisterMBean was called
Next patch will use different ObjectName implementation that will
use less memory and this patch is crucial because without it every ObjectName
is transformed with ObjectName.getInstance which turns the object into
a regular ObjectName.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
When a node is part of a cluster but is down (like in the situation where
a cluster is taken down and up again but not all nodes are up). There is
no application_state information for that node.
This patch check that the information exists before using it to prevent
null pointer exception.
After this patch, a call to nodetool gossipinfo would return the
available information without failing.
See scylladb/scylla#3330
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Message-Id: <20180329115345.29357-1-amnon@scylladb.com>
Refs scylladb/scylla#2340 (trunk/1.7)
Must proxy "register" call, otherwise unregistration of mbeans
will instead try to double-register. Code for this somehow fell away.
Message-Id: <1494417610-9720-1-git-send-email-calle@scylladb.com>
nodetool from scylla-tools-java still uses it. Currently `nodetool compact` fails like this:
error: forceKeyspaceCompaction(java.lang.String, [Ljava.lang.String;)
-- StackTrace --
java.lang.NoSuchMethodException: forceKeyspaceCompaction(java.lang.String, [Ljava.lang.String;)
Fixesscylladb/scylla#2261
Probably broken by 3e146845b4
Message-Id: <1491470483-6147-1-git-send-email-tgrabiec@scylladb.com>
This patch fixes two issues with the histogram implementation:
* Need to all update before trying to read values from the histogram.
* The histogram values return from the API in microseconds and not nano.
See Scylladb/scylla#2155
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Message-Id: <20170315130725.22261-1-amnon@scylladb.com>
This patch change the sorting of tokensEndpointMap so it will use the
order returned by the API.
See Scylladb/scylla#1945
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Note: c3 adds configurable size threshold counting of messages sent,
dividing info "large"/"small" partitions (+gossiper). Message bulk
queries in v3 mbean reflects this.
Scylla does not (yet?) have such a threshold divider, so this is
highly incomplete and just delegates to old apis that "sort-of" fit.
With the change to APITimer there is no longer a need to periodically
pull the API.
The verb will be register on the object initialization and will be
updated whenever they are been used.
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
With the change in the meter implementation, retrieving a non existing
metrics would take time.
For this, the CacheService would mark caches that are not supported with
null url, so the metrics will be register but will return 0 for all
request (instead of going to the API that will return 0).
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
CacheMetrics is a general counter that is used for all possible caches.
For caches that we do not support, there is no need to go and fetch
their values.
When moving to the APITimer each such request will take longer (the
value will not be available as it use it) now it will be possible to
supply a null as a url which would cause to the metrics to return 0 for
all counters.
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
The APITimer uses a different endpoint not to break existing API.
The addNano functionality was removed as all of the values are updated
from the APi.
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
