We cannot detect which encoding it is, if it is a non-unicode one.
I also think that will be OK in most situations because users will normally operate their client machine with LOCALE that matches the MySQL charset they use for storage (spanish/latin1 ; czech/latin2 ; russian/cp1250 etc.).
When is selected no CHARSET parameter is used in the LOAD DATA statement.
This is OK in server 5.0.51a, but not in 5.0.45. The last row does not come in the result with 5.0.45. For server versions with that bug we will fail to sync integer values having upper limit value. For an 'ordinary' INT this is 2147483647. There is a similar issue with BIGINT, possibly SMALLINT/TINYINT too. What values are affected for each yet not sure. What effect specifying UNSIGNED may have is not yet clear either. This we will have to analyze too!
This code is very complex (recursive functions calling another) and complicated to debug. And when there are server bugs as well that should be taken into acoount it is ververyvery complex …
ok, we could create a folder tree …favorites … (and a ..favorites ) and only use ALL and connection specific folder 'sub-tree'. Correct that it need not be as big a implementation change as I thought in the first place.
First you write server version 5.0.51 and next you write server version 5.0.37. Do you have more servers?
CHARSET parameter in LOAD DATA is supported from server version 5.0.38 only! For that reason that option is disabled in SQLyog if server version is lower (server would return an error!9.
Actually we tested on an old machine that had an old server version and that is why we thought tehre was a bug. Now it does not seem so.
It works for us with Win 2k and MySQL 5.0.45.
Can you reproduce with Win2K and servers >= 5.0.38? There may be some bug/regression in the 5.0.51 server then!
I hope this that be fixed (along with the other remainig 2-3 reported issues with 6.16RC2) in 1-2 days.
Now I think you can 'workaround' like this:
* just do the sync.
* use ALTER TABLE to define the index length to columns on . If no data are on it will simply change the table definition. Even if data are there existing indexes will be truncated to the length specified (and you just specify the same as originally on).
I think what your refer to as 'the index column prefix' the index length as here specified as (80)? This is the the only difference I can see!
Bug confirmed with 6.16 RC2. Actually we fixed the same issue in 'duplicate table' in version 6.04 – so we should have found this too at that occasion!
Actually adding a few rows with very large numbers (close to max. value for an BIGINT UNSIGNED) to the 20K row subset you provided will fail. This we should have observed in the first place!
A brief technical explanation: we locate rows that differ on either host calculating checksums recursively. First we take a large 'chunk' of rows. If there is difference we repeat recursively untill individual rows have been identified (if 'chunks' are identical we proceed to next chunk). We apply some mathematics to find the optimal chunk-size for every iteration.
There must be something wrong with that mathematics! We thought we had ensured that no overflow/truncation would take place (we changed the mathematics slightly and used 'wider' datatypes internally from last patch). But this is probably still the issue.
I can only confirm today – I cannot debug it from here. Our office timezone is 4,5 hours early as compared to CET and 10.5 hours early as compared to US Western Time, so everybody has left office. They will work on it before you and I get up!