afedulov commented on a change in pull request #11195: [FLINK-16222][runtime]
Use plugins mechanism for initializing MetricReporters
URL: https://github.com/apache/flink/pull/11195#discussion_r393580453
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/metrics/ReporterSetup.java
##########
@@ -178,36 +164,96 @@ private static ReporterSetup createReporterSetup(String
reporterName, MetricConf
metricReporterOptional.ifPresent(reporter -> {
MetricConfig metricConfig = new
MetricConfig();
reporterConfig.addAllToProperties(metricConfig);
-
-
reporterArguments.add(createReporterSetup(reporterName, metricConfig,
reporter));
+
reporterSetups.add(createReporterSetup(reporterName, metricConfig, reporter));
});
- }
- catch (Throwable t) {
+ } catch (Throwable t) {
LOG.error("Could not instantiate metrics
reporter {}. Metrics might not be exposed/reported.", reporterName, t);
}
}
- return reporterArguments;
+ return reporterSetups;
}
- private static Map<String, MetricReporterFactory>
loadReporterFactories() {
- final ServiceLoader<MetricReporterFactory> serviceLoader =
ServiceLoader.load(MetricReporterFactory.class);
+ private static List<Tuple2<String, Configuration>>
loadReporterConfigurations(Configuration configuration, Set<String>
namedReporters) {
+ final List<Tuple2<String, Configuration>>
reporterConfigurations = new ArrayList<>(namedReporters.size());
+
+ for (String namedReporter: namedReporters) {
+ DelegatingConfiguration delegatingConfiguration = new
DelegatingConfiguration(
+ configuration,
+ ConfigConstants.METRICS_REPORTER_PREFIX +
namedReporter + '.');
+ reporterConfigurations.add(Tuple2.of(namedReporter,
delegatingConfiguration));
+ }
+ return reporterConfigurations;
+ }
+
+ private static Set<String>
findEnabledReportersInConfiguration(Configuration configuration, String
includedReportersString) {
+ Set<String> includedReporters =
reporterListPattern.splitAsStream(includedReportersString)
+ .filter(r -> !r.isEmpty()) // splitting an empty string
results in an empty string on jdk9+
+ .collect(Collectors.toSet());
+
+ // use a TreeSet to make the reporter order deterministic,
which is useful for testing
+ Set<String> namedOrderedReporters = new
TreeSet<>(String::compareTo);
+
+ // scan entire configuration for keys starting with
METRICS_REPORTER_PREFIX and determine the set of enabled reporters
+ for (String key : configuration.keySet()) {
+ if
(key.startsWith(ConfigConstants.METRICS_REPORTER_PREFIX)) {
+ Matcher matcher =
reporterClassPattern.matcher(key);
+ if (matcher.matches()) {
+ String reporterName = matcher.group(1);
+ if (includedReporters.isEmpty() ||
includedReporters.contains(reporterName)) {
+ if
(namedOrderedReporters.contains(reporterName)) {
+ LOG.warn("Duplicate
class configuration detected for reporter {}.", reporterName);
+ } else {
+
namedOrderedReporters.add(reporterName);
+ }
+ } else {
+ LOG.info("Excluding reporter
{}, not configured in reporter list ({}).", reporterName,
includedReportersString);
+ }
+ }
+ }
+ }
+ return namedOrderedReporters;
+ }
+
+ private static Map<String, MetricReporterFactory>
loadAvailableReporterFactories(PluginManager pluginManager) {
final Map<String, MetricReporterFactory> reporterFactories =
new HashMap<>(2);
- final Iterator<MetricReporterFactory> factoryIterator =
serviceLoader.iterator();
+ final Iterator<MetricReporterFactory> factoryIterator =
getAllReporterFactories(pluginManager);
+ LOG.debug("All available factories (from both SPIs and
Plugins):");
+ getAllReporterFactories(pluginManager).forEachRemaining(i ->
LOG.debug(i.toString()));
// do not use streams or for-each loops here because they do
not allow catching individual ServiceConfigurationErrors
// such an error might be caused if the META-INF/services
contains an entry to a non-existing factory class
while (factoryIterator.hasNext()) {
try {
MetricReporterFactory factory =
factoryIterator.next();
-
reporterFactories.put(factory.getClass().getName(), factory);
+ String factoryClassName =
factory.getClass().getName();
+ MetricReporterFactory existingFactory =
reporterFactories.get(factoryClassName);
+ if (existingFactory == null){
+ reporterFactories.put(factoryClassName,
factory);
+ LOG.warn(new
File(factory.getClass().getProtectionDomain().getCodeSource().getLocation()
+ .toURI()).getCanonicalPath());
+ } else {
+ //TODO: use path information below,
when Plugin Classloader stops always prioritizing factories from /lib
+// String jarPath1 = new
File(existingFactory.getClass().getProtectionDomain().getCodeSource().getLocation()
+// .toURI()).getCanonicalPath();
+// String jarPath2 = new
File(factory.getClass().getProtectionDomain().getCodeSource().getLocation()
+// .toURI()).getCanonicalPath();
+// LOG.warn("Multiple implementations of
the same reporter were found: \n {} and \n{}", jarPath1, jarPath2);
+ LOG.warn("Multiple implementations of
the same reporter were found in 'lib' and 'plugins' directories for {}. It is
recommended to remove redundant reporter JARs to resolve used versions'
ambiguity.", factoryClassName);
Review comment:
Strictly speaking it then should be 'and/or', because 'or' could mean that
there are either two jars both either in 'lib' or 'plugins'. I think and was
more fitting, but I am also ok with and/or
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
