it seems you are using 2.4.0. It does not seem related, but this fix
here would be important to have it on your system:

commit 6b1abd1aadc2d097e3baefeb312c8e68092876ba
Author: Clebert Suconic <[email protected]>
Date:   Sun Aug 26 15:55:56 2018 -0400
    ARTEMIS-2053 avoiding data loss after compacting



However, let me explain you how the record scanning works:

the format for the data is:

Roughly
JOURNAL-RECORD-TYPE (byte)
FILE-ID (int)
compact-count(byte)
recordID(long)
recordSize, from persisters (int)
userRecordType
total-record-size



When we recycle a file, we simply change the fileID on the header, and
when we load the file, the scan is done by matching the record-type
and at the end of the record the total-record-size has to match the
record-type.

I did this to avoid filling up the file with zeros, which at the time
was a costly operation (I wrote this when disks were still mechanical
at the time), but that's still a costly operation.

So, to wrongly trick the scan you will need a byte record, matching
the fileID at the next int, with the recordsize and total-record-size
matching each other.


Perhaps the loading is skipping the verification on total-record-size
and that snicked an invalid record?



Or perhaps the fact that you missed the commit I mentioned caused you an issue?



On Mon, Jul 1, 2019 at 5:57 AM yw yw <[email protected]> wrote:
>
> Hi, All
>
>
> Yesterday our cluster experienced a sudden loss of power. When we started
> broker after power brought back, exception occurred:
>
>
> The exception showed the userRecordType loaded was illegal. The operation
> team deleted data journals and broker started successfully.
>
>
> It was a pity we didn't backup the problematic journal files. We checked
> dmesg command output, no disk errors. SMART tests on disk also showed disk
> not broken. Then we digged into code(JournalImpl::readJournalFile) and
> tried to find something. We have two doubts with the code.
>
>
> First doubt:
>
> The comment says "I - We scan for any valid record on the file. If a hole
> happened on the middle of the file we keep looking until all the
> possibilities are gone".
>
> Considering we're appending journal file and fileId is strictly increasing,
> so we can just skip the whole file if the fileId of record is not equal to
> file id. IMO the rest records in the file are the same, no need to read
> them. Should we keep looking all the possibilities, is there a
> possibility(very low one) that we just assemble a record of which fileId,
> recordType, checkSize all qualifies but actually does not exist?
>
> Our second one:
>
> In the case of power outage where part of record is written into disk, e.g.
> recordyType,fileId is successfully written, we may read the old record
> though fileId is latest?
>
> Can anyone shed some lights on this please? Thanks.



-- 
Clebert Suconic

Reply via email to