Commit Graph

65 Commits

Author SHA1 Message Date
Oleg Zhurakousky 37635232c7 NIFI-1218 upgraded Kafka to 0.9.0.0 client API Tested and validated that it is still compatible with 0.8.* Kafka brokers
Signed-off-by: jpercivall <joepercivall@yahoo.com>
2015-12-16 14:25:28 -05:00
joewitt d755e43ec8 NIFI-1122_nifi-0.4.0-RC2prepare for next development iteration 2015-12-08 13:00:10 -05:00
joewitt b66c029090 NIFI-1122_nifi-0.4.0-RC2prepare release nifi-0.4.0-RC2 2015-12-08 12:59:59 -05:00
Mark Payne 4fa2a713a2 NIFI-1192: Removed some additional white space 2015-11-25 12:57:06 -05:00
Oleg Zhurakousky d949ee1a1e NIFI-1192 added support for dynamic properties to GetKafka
Due to the fact that current component uses artificial names for properties set via UI and then maps those properties to the actual names used by Kafka, we can not rely on NiFi UI to display an error if user attempts to set a dynamic property which will eventually map to the same Kafka property. So, I’ve decided that any dynamic property will simply override an existing property with WARNING message displayed. It is actually consistent with how Kafka does it and displayed the overrides in the console. Updated the relevant annotation description.
It is also worth to mentioned that current code was using an old property from Kafka 0.7 (“zk.connectiontimeout.ms”) which is no longer present in Kafka 0.8 (WARN Timer-Driven Process Thread-7 utils.VerifiableProperties:83 - Property zk.connectiontimeout.ms is not valid). The add/override strategy would provide for more flexibility when dealing with Kafka volatile configuration until things will settle down and we can get some sensible defaults in place.

While doing it addressed the following issues that were discovered while making modification and testing:
ISSUE: When GetKafka started and there are no messages in Kafka topic the onTrigger(..) method would block due to the fact that Kafka’s ConsumerIterator.hasNext() blocks. When attempt was made to stop GetKafka would stops successfully due to the interrupt. However in UI it would appear as ERROR based on the fact that InterruptException was not handled.
RESOLUTION: After discussing it with @markap14 the the general desire is to let the task exit as quick as possible and that the whole thread maintenance logic was there initially due to the fact that there was no way to tell Kafka consumer to return immediately if there are no events. In this patch we are now using ‘consumer.timeout.ms’ property of Kafka and setting its value to 1 millisecond (default is -1 - always block infinitely). This ensures that tasks that attempted to read an empty topic will exit immediately just to be rescheduled by NiFi based on user configurations.

ISSUE:  Kafka would not release FlowFile with events if it didn’t have enough to complete the batch since it would block waiting for more messages (based on the blocking issue described above).
RESOLUTION: The invocation of hasNext() results in Kafka’s ConsumerTimeoutException which is handled in the catch block where the FlowFile with partial batch will be released to success. Not sure if we need to put a WARN message. In fact in my opinion we should not as it may create unnecessary confusion.

ISSUE: When configuring a consumer for topic and specifying multiple concurrent consumers in ‘topicCountMap’ based on 'context.getMaxConcurrentTasks()’ each consumer would bind to a topic partition. If you have less partitions then the value returned by 'context.getMaxConcurrentTasks()’ you would essentially allocate Kafka resources that would never get a chance to receive a single message  (see more here https://cwiki.apache.org/confluence/display/KAFKA/Consumer+Group+Example).
RESOLUTION: Logic was added to determine the amount of partitions for a topic and in the event where 'context.getMaxConcurrentTasks()’ value is greater than the amount of partitions, the partition count will be used to when creating ‘topicCountMap’ and WARNING message will be displayed)see code). Unfortunately we can’t do anything with the actual tasks, but based on current state of the code they will exit immediately just to be rescheduled where the process will repeat. NOTE: That is not ideal as it will be rescheduling tasks that will never have a chance to do anything, but at least it could be fixed on the user side after reading the warning message.

NIFI-1192 added dynamic properties support for PutKafka

NIFI-1192 polishing

NIFI-1192 polished and addressed PR comments
2015-11-24 12:14:36 -05:00
Naveen Madhire 03a54bf2d5 Changes after review 2015-11-16 13:32:17 -06:00
Naveen Madhire b954ca620e NIFI-1146 Allow GetKafka to be configured with auto.offset.reset to largest or smallest 2015-11-16 11:59:52 -06:00
Mark Payne 22de23baa6 NIFI-1097: Rewrite PutKafka to use the new producer api 2015-11-13 10:46:46 -05:00
joewitt 99629646fe NIFI-1122 moved to 0.4.0-SNAPSHOT 2015-11-06 23:41:15 -05:00
Mark Payne 9515b74607 NIFI-1088: Ensure that FlowFile is penalized before routing to failure 2015-10-30 14:25:27 -04:00
Mark Payne 4afd8f88f8 NIFI-810: Created RequiresInput annotation and ensure that processors are invalid if connections do not agree 2015-10-07 17:26:14 -04:00
Randy Gelhausen 31fba6b333 Bumped nifi-kafka-processors Kafka version
This closes #99.

Signed-off-by: Aldrin Piri <aldrin@apache.org>
2015-10-05 10:39:22 -04:00
Matt Gilman ded74ec94c NIFI-876 prepare for next development iteration 2015-09-14 21:48:11 -04:00
Matt Gilman 2ec735e350 NIFI-876 prepare release nifi-0.3.0-RC1 2015-09-14 21:48:00 -04:00
joewitt aa99884782 NIFI-850 removed nifi parent, updated nifi pom, moved all nifi subdirs up one level, fixed readme. 2015-08-15 13:12:22 -04:00