Thanks to both of you. Some quick points: I'd expect there to be backpressure from the producer if the broker is busy ie. the broker would not respond to the console producer if the broker was too busy accept more messages, and the producer would hang on the socket. Alternatively I'd hope the console producer would have the sense to back off and retry but clearly(?) not. This behaviour is actually relevant to my old job so I need to know more.
Perhaps the timeout mentioned in the error msg can just be upped? *Is* the claimed timeout relevant? > Batch containing 8 record(s) expired due to timeout while requesting metadata > from brokers for big_ptns1_repl1_nozip-0 Why is the producer expiring records? But I'm surprised this happened because my setup is one machine with everything running on it. No network. Also Kafka writes to the disk without an fsync (or its equivalent on windows) which means it just gets cached in ram before being lazily written to disk, and I've got plenty of ram - 16GB ram vs 5GB of input file. Kafka adds its overhead so it grows to ~8GB but still, it need not hit disk at all (and the file goes into the windows memory, not java's). Maybe it is GC holding things up but I dunno, GC even for a second or two should not cause a socket failure, just delay the read, though I'm not an expert on this *at all*. I'll go over the answers tomorrow more carefully but thanks anyway! cheers jan On 18/04/2017, Serega Sheypak <serega.shey...@gmail.com> wrote: >> err, isn't it supposed to? Isn't the loss of data a very serious error? > Kafka can't fix networking issues like latencies, blinking, unavailability > or any other weird stuff. Kafka promises you to persist data if data > reaches Kafka. Data delivery responsibility to kafka is on your side. You > fail to do it according to logs. > > 0.02% not 2% > You should check broker logs to figure out what went wrong. All things > happen on one machine as far as I understand. Maybe your brokers don't have > enough mem and they stuck because of GC and don't respond to producer. > Async producer fails to send data. That is why you observe data loss on > consumer side. > > > 2017-04-18 23:32 GMT+02:00 jan <rtm4...@googlemail.com>: > >> Hi Serega, >> >> > data didn't reach producer. So why should data appear in consumer? >> >> err, isn't it supposed to? Isn't the loss of data a very serious error? >> >> > loss rate is more or less similar [...] Not so bad. >> >> That made me laugh at least. Is kafka intended to be a reliable >> message delivery system, or is a 2% data loss officially acceptable? >> >> I've been reading the other threads and one says windows is really not >> supported, and certainly not for production. Perhaps that's the root >> of it. Well I'm hoping to try it on linux shortly so I'll see if I can >> replicate the issue but I would like to know whether it *should* work >> in windows. >> >> cheers >> >> jan >> >> On 18/04/2017, Serega Sheypak <serega.shey...@gmail.com> wrote: >> > Hi, >> > >> > [2017-04-17 18:14:05,868] ERROR Error when sending message to topic >> > big_ptns1_repl1_nozip with key: null, value: 55 bytes with error: >> > (org.apache.kafka.clients. >> > producer.internals.ErrorLoggingCallback) >> > org.apache.kafka.common.errors.TimeoutException: Batch containing 8 >> > record(s) expired due to timeout while requesting metadata from >> > brokers for big_ptns1_repl1_nozip-0 >> > >> > data didn't reach producer. So why should data appear in consumer? >> > loss rate is more or less similar : 0.02 (130k / 5400mb) ~ 0.03% (150mb >> > / >> > 5000gb) Not so bad. >> > >> > >> > 2017-04-18 21:46 GMT+02:00 jan <rtm4...@googlemail.com>: >> > >> >> Hi all, I'm something of a kafka n00b. >> >> I posted the following in the google newsgroup, haven't had a reply >> >> or even a single read so I'll try here. My original msg, slightly >> >> edited, was: >> >> >> >> ---- >> >> >> >> (windows 2K8R2 fully patched, 16GB ram, fairly modern dual core xeon >> >> server, latest version of java) >> >> >> >> I've spent several days trying to sort out unexpected behaviour >> >> involving kafka and the kafka console producer and consumer. >> >> >> >> If I set the console produced and console consumer to look at the >> >> same topic then I can type lines into the producer window and see them >> >> appear in the consumer window, so it works. >> >> >> >> If I try to pipe in large amounts of data to the producer, some gets >> >> lost and the producer reports errors eg. >> >> >> >> [2017-04-17 18:14:05,868] ERROR Error when sending message to topic >> >> big_ptns1_repl1_nozip with key: null, value: 55 bytes with error: >> >> (org.apache.kafka.clients. >> >> producer.internals.ErrorLoggingCallback) >> >> org.apache.kafka.common.errors.TimeoutException: Batch containing 8 >> >> record(s) expired due to timeout while requesting metadata from >> >> brokers for big_ptns1_repl1_nozip-0 >> >> >> >> I'm using as input a file either shakespeare's full works (about 5.4 >> >> meg ascii), or a much larger file of shakespear's full works >> >> replicated 900 times to make it about 5GB. Lines are ascii and short, >> >> and each line should be a single record when read in by the console >> >> producer. I need to do some benchmarking on time and space and this >> >> was my first try. >> >> >> >> As mentioned, data gets lost. I presume it is expected that any data >> >> we pipe into the producer should arrive in the consumer, so if I do >> >> this in one windows console: >> >> >> >> kafka-console-consumer.bat --bootstrap-server localhost:9092 --topic >> >> big_ptns1_repl1_nozip --zookeeper localhost:2181 > >> >> F:\Users\me\Desktop\shakespear\single_all_shakespear_OUT.txt >> >> >> >> and this in another: >> >> >> >> kafka-console-producer.bat --broker-list localhost:9092 --topic >> >> big_ptns1_repl1_nozip < >> >> F:\Users\me\Desktop\shakespear\complete_works_no_bare_lines.txt >> >> >> >> then the output file "single_all_shakespear_OUT.txt" should be >> >> identical to the input file "complete_works_no_bare_lines.txt" except >> >> it's not. For the complete works (sabout 5.4 meg uncompressed) I lost >> >> about 130K in the output. >> >> For the replicated shakespeare, which is about 5GB, I lost about 150 >> meg. >> >> >> >> This can't be right surely and it's repeatable but happens at >> >> different places in the file when errors start to be produced, it >> >> seems. >> >> >> >> I've done this using all 3 versions of kafak in the 0.10.x.y branch >> >> and I get the same problem (the above commands were using the 0.10.0.0 >> >> branch so they look a little obsolete but they are right for that >> >> branch I think). It's cost me some days. >> >> So, am I making a mistake, if so what? >> >> >> >> thanks >> >> >> >> jan >> >> >> > >> >
