Re: PME speedup #2, TX recovery delay elimination.
Anton, > Since this fix made to speedup pme-free switch which prohibits the merges... Sounds as a bit of knowledge that I missed. Good to understand it now. > The "magic numbers" are always the "magic numbers" :) > We must get rid of them to see problems covered by them. Let's try =) чт, 26 дек. 2019 г. в 15:09, Anton Vinogradov : > > Ivan, > > The "magic numbers" are always the "magic numbers" :) > We must get rid of them to see problems covered by them. > > >> Was there any > >> performance measurements for such multiple left nodes cases? > Since this fix made to speedup pme-free switch which prohibits the merges, > the answer is "no". > > BTW, the fix was merged to master. > > On Thu, Dec 26, 2019 at 2:21 PM Ivan Pavlukhin wrote: > > > Anton, > > > > Thank you for your efforts! And sorry for a late reply. > > > > I am a little bit familiar with tx recovery. I personally like the > > idea of removing such "magic" logic from the code. I think a proper > > way is either justify and sustain (tests, documentation) some behavior > > or get rid of it. > > > > Regarding a delay before tx recovery. My understanding was that it > > might be useful when multiple (client) nodes leaves almost at the same > > time (perhaps due to some network connectivity issues). With a delay > > recovering multiple failed nodes will be grouped into one recovery > > round (+PME). Correct me if my understanding is wrong. Was there any > > performance measurements for such multiple left nodes cases? > > > > вт, 24 дек. 2019 г. в 16:22, Anton Vinogradov : > > > > > > Rechecked TC two more times. > > > Going to merge to master in case no objections here. > > > > > > On Mon, Dec 23, 2019 at 1:44 PM Anton Vinogradov wrote: > > > > > > > Igniters, > > > > > > > > One more PME optimization ready to be reviewed. > > > > I found a strange tx recovery delay caused by > > IGNITE_TX_SALVAGE_TIMEOUT. > > > > I've checked the code and tests and found no reason to delay recovery. > > > > > > > > So, the issue [1] is ready to be reviewed. > > > > > > > > Improvement checked with benchmark [2] and fix, obviously, 100 ms > > faster > > > > :) > > > > > > > > [1] https://issues.apache.org/jira/browse/IGNITE-12272 > > > > [2] > > > > > > https://github.com/anton-vinogradov/ignite/commit/f8c27253b0ecfe7381418f505aafe942efe5a0a8 > > > > > > > > > > > > -- > > Best regards, > > Ivan Pavlukhin > > -- Best regards, Ivan Pavlukhin
Re: PME speedup #2, TX recovery delay elimination.
Ivan, The "magic numbers" are always the "magic numbers" :) We must get rid of them to see problems covered by them. >> Was there any >> performance measurements for such multiple left nodes cases? Since this fix made to speedup pme-free switch which prohibits the merges, the answer is "no". BTW, the fix was merged to master. On Thu, Dec 26, 2019 at 2:21 PM Ivan Pavlukhin wrote: > Anton, > > Thank you for your efforts! And sorry for a late reply. > > I am a little bit familiar with tx recovery. I personally like the > idea of removing such "magic" logic from the code. I think a proper > way is either justify and sustain (tests, documentation) some behavior > or get rid of it. > > Regarding a delay before tx recovery. My understanding was that it > might be useful when multiple (client) nodes leaves almost at the same > time (perhaps due to some network connectivity issues). With a delay > recovering multiple failed nodes will be grouped into one recovery > round (+PME). Correct me if my understanding is wrong. Was there any > performance measurements for such multiple left nodes cases? > > вт, 24 дек. 2019 г. в 16:22, Anton Vinogradov : > > > > Rechecked TC two more times. > > Going to merge to master in case no objections here. > > > > On Mon, Dec 23, 2019 at 1:44 PM Anton Vinogradov wrote: > > > > > Igniters, > > > > > > One more PME optimization ready to be reviewed. > > > I found a strange tx recovery delay caused by > IGNITE_TX_SALVAGE_TIMEOUT. > > > I've checked the code and tests and found no reason to delay recovery. > > > > > > So, the issue [1] is ready to be reviewed. > > > > > > Improvement checked with benchmark [2] and fix, obviously, 100 ms > faster > > > :) > > > > > > [1] https://issues.apache.org/jira/browse/IGNITE-12272 > > > [2] > > > > https://github.com/anton-vinogradov/ignite/commit/f8c27253b0ecfe7381418f505aafe942efe5a0a8 > > > > > > > -- > Best regards, > Ivan Pavlukhin >
Re: PME speedup #2, TX recovery delay elimination.
Anton, Thank you for your efforts! And sorry for a late reply. I am a little bit familiar with tx recovery. I personally like the idea of removing such "magic" logic from the code. I think a proper way is either justify and sustain (tests, documentation) some behavior or get rid of it. Regarding a delay before tx recovery. My understanding was that it might be useful when multiple (client) nodes leaves almost at the same time (perhaps due to some network connectivity issues). With a delay recovering multiple failed nodes will be grouped into one recovery round (+PME). Correct me if my understanding is wrong. Was there any performance measurements for such multiple left nodes cases? вт, 24 дек. 2019 г. в 16:22, Anton Vinogradov : > > Rechecked TC two more times. > Going to merge to master in case no objections here. > > On Mon, Dec 23, 2019 at 1:44 PM Anton Vinogradov wrote: > > > Igniters, > > > > One more PME optimization ready to be reviewed. > > I found a strange tx recovery delay caused by IGNITE_TX_SALVAGE_TIMEOUT. > > I've checked the code and tests and found no reason to delay recovery. > > > > So, the issue [1] is ready to be reviewed. > > > > Improvement checked with benchmark [2] and fix, obviously, 100 ms faster > > :) > > > > [1] https://issues.apache.org/jira/browse/IGNITE-12272 > > [2] > > https://github.com/anton-vinogradov/ignite/commit/f8c27253b0ecfe7381418f505aafe942efe5a0a8 > > -- Best regards, Ivan Pavlukhin
Re: PME speedup #2, TX recovery delay elimination.
Rechecked TC two more times. Going to merge to master in case no objections here. On Mon, Dec 23, 2019 at 1:44 PM Anton Vinogradov wrote: > Igniters, > > One more PME optimization ready to be reviewed. > I found a strange tx recovery delay caused by IGNITE_TX_SALVAGE_TIMEOUT. > I've checked the code and tests and found no reason to delay recovery. > > So, the issue [1] is ready to be reviewed. > > Improvement checked with benchmark [2] and fix, obviously, 100 ms faster > :) > > [1] https://issues.apache.org/jira/browse/IGNITE-12272 > [2] > https://github.com/anton-vinogradov/ignite/commit/f8c27253b0ecfe7381418f505aafe942efe5a0a8 >
