Where communities thrive

  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
Repo info
    Paweł Niedzielski

    Initial search pointed to a couple of issues in python library (which uses librdkafka, just like phprdkafka does).

    Could you please provide an excerpt from your code near commit call, check if the issue originates around that location @ruschoni02 ?

    Also please check which versions of phprdkafka & librdkafka you're using. From phpinfo() or php -i | grep -C 1 kafka
    Антон Колинько

    Hey guys, would like to get some recommendation from you. I am new to kafka, so don't yet know some best practices. We have a producers written on different languages and it's already running in prod. From php side I need to write a consumers that will run long term via supervisor in GCP.

    What I want to ask is what are the best settings for running kafka php consumer in long term? What settings you can recommend to change and to what value?

    Currently I ended up with the following:
    General config:

    $conf = new \RdKafka\Conf();
    $conf->set('compression.codec', 'gzip');
    $conf->set('', '86400000'); <- is that one ok?

    Topic config:

    $topicConf = new \RdKafka\TopicConf();
    $topicConf->set('auto.commit.enable', 'false');
    $topicConf->set('', 5000); <- what value is best here?
    $topicConf->set('', 'broker'); <- is it good to use or "file" should be used? 
    $topicConf->set('auto.offset.reset', 'latest');

    This is how I consume messages (simplified):

    $topic->consumeStart($partition, RD_KAFKA_OFFSET_STORED);
    while (true) {
        $message = $topic->consume($partition, 1000); <- what time it's better to set here?
        $messageErr = is_object($message) ? $message->err : RD_KAFKA_RESP_ERR_UNKNOWN;
        switch ($messageErr) {
            case RD_KAFKA_RESP_ERR_NO_ERROR:
                echo $message->payload . PHP_EOL;
        // After successfully consuming the message, schedule offset store.
        // Offset is actually committed after '' milliseconds.
        $topic->offsetStore($message->partition, $message->offset);

    I highlight the important settings with arrow <-. Can someone please advise?

    Paweł Niedzielski

    @AntKolinko_twitter - this value decides how often client will send a request (small) to commit his offset. Your setting is ok, since afaik by default phprdkafka does that in an async manner, not blocking. You might want to disable (EDIT: Which you did, see below) automatic commit if your throughput is extremely high (like 10k messages per sec) and you need to squeeze those additional ms's)
    auto.commit.enable - actually, disregard the above. You've disabled it. - keep it on broker. It's default for newer Kafka servers. The difference is that brokers store offsets in a dedicated topic (and sync) instead of keeping it in a file. should not be set that high afaik. Keeping it on default 300000 should be ok. It shouldn't have too much of an impact on your application, unless there is more than 30s interval between your consume calls (Kafka will assume client is dead).

    I'm not 100% sure about, since I'm a bit confused about it. Especially since librdkafka introduced some changes in 1.0.0 regarding it afaik.

    Антон Колинько

    @steveb_gitlab Thank you very much for response! Did I get it right that since auto.commit.enable = false I can remove this setting at all as commit happens manually?

    About this one didn't understand how that works. Does it mean that if producers send 1 message per 10 minute and I have value set to 300000ms (which is 6 minutes) then my consumer will be considered as dead?

    Paweł Niedzielski
    @steveb_gitlab regulates how often your producers/consumers are expected to "poll". Poll is done when consumer calls consume or producer calls poll (in phprdkafka this equals waiting for message to actually be posted to kafka). If your process does not communicate in time it is considered "dead" by Kafka.

    This usually means that producer get disconnected, and consumer groups are rebalanced.

    If I remember correctly, producer will automatically reconnect if needed, but I'm not 100% sure
    and since your commiting manually you can safely omit, as it should do nothing (unless there is some bug :P)
    see (since most options are passed directly to the underlying librdkafka library)

    Maximum allowed time between calls to consume messages (e.g., rd_kafka_consumer_poll()) for high-level consumers. If this interval is exceeded the consumer is considered failed and the group will rebalance in order to reassign the partitions to another consumer group member. Warning: Offset commits may be not possible at this point. Note: It is recommended to set

    Actually should only affect consumers

    Антон Колинько

    ok, thanks. Will remove, so default time is used and remove this one

    What do you think about $message = $topic->consume($partition, 1000); <- what time it's better to set here?

    Paweł Niedzielski
    @AntKolinko_twitter doesn't matter too much. This timeout is the maximum amount of time that this one particular call can wait. You should be able to safely increase it to like 10 seconds. It's use-case is for application that run with an event loop, so they can transfer control back to the loop. Since you're in a while(true) loop it doesn't matter that much.
    You're never really exiting the loop.
    How to run kafka on different machines ?
    When i stop any one server then another one is able to send the details .
    Paweł Niedzielski
    @abdo1 I don't really understand the question. Are you trying to run multiple Phprdkafka clients connected to the same topic?
    Wesley Willians
    Hello Guys, I'm working with kafka and PHP and I need to make a sync request. I'm trying to use Request/Reply pattern, but I'm facing a problem. My consumer always have a delay to be connected so that my producer already sent the msg and it cant be read unless I read all data from the topic.. any insight? The idea is a kind of PubSub.
    btw, how can I get a current offset of a topic?
    Kamil Hurajt
    Hi Guys, I'm first time trying to use kafka as FIFO queue, as in company it's implemented. But I'm facing issue that by using High-Level consuming the service is waiting for mesage for 40 seconds!!! What I'm doing wrong ?
    Amy Luo
    Is there a way I can synchronously produce messages in php-rdkafka library?
    Paweł Niedzielski
    @amyluo just flush it immediately :D
    Aleksandar Aramov
    Hello guys, RdKafka\Conf::setDefaultTopicConf This function has been DEPRECATED. Where I should put Topic config? Inside RdKafka\Conf::set, should I use RdKafka\TopicConf class ? Thanks in advance.
    Paweł Niedzielski
    @alex687 practically all configuration settings that have been available in TopicConf can be set in Conf object directly. Those that cannot can still be passed as TopicConf object to those classes that can receive it as argument (which is in practice all places that previously used it)
    Aleksandar Aramov
    @steveb_gitlab Thanks, I was wondering, because when I dump config ( config->dump()) it topic configs were not there.
    Paweł Niedzielski
    @alex687 actually, there is an issue for something like that
    I'm not sure it applies
    late to the party, but yeah @steveb_gitlab is right i opened this issue because of this :)
    Zakhar Morozov
    Hello. Is there any way to use truststore with SASL_SSL? I'm getting the error Java TrustStores are not supported, use '' and a certificate file instead. when i use ssl.truststore.location
    Paweł Niedzielski
    @fortael according to this comment from 2017 it's not
    and it's not really an issue with phprdkafka (or librdkafka actually)
    Zakhar Morozov
    I'm not quite understood you. Our java team use it in that configuration. So i expect php-extension able to do the same. I tried to use .pem file but the i got Java JAAS configuration is not supported too.
    Paweł Niedzielski
    @fortael Java client is not the same as librdkafka (C library for Kafka client for all other languages - node, python, or in this case php)
    librdkafka itself does not support Java Truststores as the exception is telling you
    Zakhar Morozov
    Got it
    Paweł Niedzielski
    Judging from the fact that the librdkafka community didn't implement it it's either proprietary and cannot be replicated to librdkafka, or it's complicated for some other reason
    but I'd say the former
    Paweł Niedzielski
    btw @fortael
    see this discussion for possible solution for you: arnaud-lb/php-rdkafka#225

    hey everyone!

    I am working on some very old legacy PHP project and I don't have much room to improve on stuff. We recently implemented sending to kafka and issue now is we have quite a lot of connections from PHP project to Kafka in TIME_WAIT.
    Unfortunately we are opening new connection for each request which is far for optimal. My questions is, is there some way to close TCP connection after flush().

    another questions is: can I somehow try utilize for producing messages?
    I think freeing the producer (setting it to null) probably will do the trick. also if you are pre librdkafka:1.0, this might too:
    from 1.x onward this is not necessary anymore
    if destroying the producer object is not an option can maybe used to achieve this too but i haven't tried it.
    thank you Nick I will try it!

    Good day everyone!
    I am trying to test out this library, but have strange behaviour on high-level consumer, default auto rebalancing in specific. I suppose there are some mistakes with Conf() settings, but have troubles with debugging on my own, can you help me out?
    We using Docker for local tests. There are eight partitions in the topic. While consuming messages with 4 sec timeout Consumer provides this info:

    [17:39:16][partition 0] Timed out [key: '' offset: 0]
    [17:39:20][partition 0] Timed out [key: '' offset: 0]
    [17:39:20][partition 4] No more messages; will wait for more [key: '' offset: 14]
    [17:39:24][partition 0] Timed out [key: '' offset: 0]
    [17:39:28][partition 0] Timed out [key: '' offset: 0]
    [17:39:32][partition 0] Timed out [key: '' offset: 0]
    [17:39:36][partition 0] Timed out [key: '' offset: 0]
    [17:39:40][partition 0] Timed out [key: '' offset: 0]
    [17:39:44][partition 0] Timed out [key: '' offset: 0]
    [17:39:48][partition 0] Timed out [key: '' offset: 0]
    [17:39:52][partition 0] Timed out [key: '' offset: 0]
    [17:39:56][partition 0] Timed out [key: '' offset: 0]
    [17:40:00][partition 0] Timed out [key: '' offset: 0]
    [17:40:04][partition 0] Timed out [key: '' offset: 0]
    [17:40:08][partition 0] Timed out [key: '' offset: 0]
    [17:40:12][partition 0] Timed out [key: '' offset: 0]
    [17:40:16][partition 0] Timed out [key: '' offset: 0]
    [17:40:20][partition 0] Timed out [key: '' offset: 0]
    [17:40:24][partition 0] Timed out [key: '' offset: 0]
    [17:40:28][partition 0] Timed out [key: '' offset: 0]
    [17:40:32][partition 0] Timed out [key: '' offset: 0]
    [17:40:36][partition 0] Timed out [key: '' offset: 0]

    Why Consumer sticks only to partition 0 with timeout error although there are other seven partitions, even with successful consuming?

    Here is current Conf() state
    $this->config = new Conf();
    $this->config->set('', env('KAFKA_REFRESH_INTERVAL_MSEC')); // 10000
    $this->config->set('', env('KAFKA_BROKER_LIST'));
    $this->config->set('auto.offset.reset', 'largest');
    $this->config->set('enable.partition.eof', 'true');
    can you elaborte a bit more, so are messages actually not being consumed or are you just wondering, after consumption, why it sticks to partition 0?
    Ok, so i checked the code, librdkafka actually returns null on timeout, so the 0 is coming from us (the extension, 99% sure), so you can ignore the partition info on this "error" since we don't have that info, i will check if we can improve this, so partition is NULL. If your concern is, that the consumer is not "jumping" around and picking up new messages from other partitions, no worries, works like a charm ;)
    PR is created ;) should be fixed soon(ish) :D

    Oh, thanks a lot! You are awesome with fast responding =)

    so are messages actually not being consumed

    Yeah, this is the case. I am 100% sure there are messages in the topic for each partition (check it with Kafdrop/Kafkamanager), but consumer skips most of them with an error, very rarely successfully consuming a message. There is a single consumer for the topic within a group:

    $this->config->set('', 'kafka_consumer_group_' . $topic);

    I add a rebalance callback as documented example stated for logging partition assign and got this:

    [10:46:41] Timed out
    [10:46:45] Timed out
    Assign: array(1) {
      object(RdKafka\TopicPartition)#483 (3) {
        string(9) "test"
    [10:46:46][partition 0] No more messages; will wait for more [key: '' offset: 22]
    [10:46:50] Timed out
    [10:46:54] Timed out
    [10:46:58] Timed out
    [10:47:02] Timed out
    [10:47:06] Timed out

    I wonder if there are problems with offsets (I use auto.offset.reset = largest).

    most certainly, earliest means: for the very first time, start at the beginning, latest means: for the very first time, start at the end
    so in your case you are only consuming new messages.