Hi, Sachin! On Aug 22, Sachin Setia wrote:
Hi Sergei!
Actually I completed the work on update and delete. Now they will use index for looking up records.
But I am thinking I have done a lot of changes in optimizer which may break it, and also there are lots of queries where my code does not work, fixing this might take a long amount of time. I am thinking of a change in my existing code :- Suppose a table t1 create table t1 (a blob, b blob, c blob, unique(a,b,c)); In current code, for query like there will a KEY with only one keypart which points to field DB_ROW_HASH_1. It was okay for normal updates, insert and delete, but in the case of where optimization I have do a lot of stuff, first to match field (like in add_key_part), then see whether all the fields in hash_str are present in where or not, then create keys by calculating hash. I do this by checking the HA_UNIQUE_HASH flag in KEY, but this also makes (I think) optimizer code bad because of too much dependence. Also I need to patch get_mm_parts and get_mm_leaf function, which I think should not be patched.
Later today I'll know exactly what you mean, when I'll finish reviewing your optimizer changes. But for now, let's say I agree on a general principle :) Optimizer is kinda complex and fragile, so it's good to avoid doing many changes in it - the effect might be difficult to predict.
I am thinking of a another approach to this problem at server level instead of having just one keypart we can have 1+3 keypart. Last three keyparts will be for field a, b, c and first one for DB_ROW_HASH_1.These will be only at server level not at storage level. key_info->key_part will point at keypart containing field a, while key_part having field DB_ROW_HASH_1 will -1 index. By this way I do not have to patch more of optimizer code. But there is one problem, what should be the length of key_part? I am thinking of it equal to field->pack_length(), this would not work because while creating keys optimizer calls get_key_image() (which is real data so can exceed pack_lenght() in case of blob), so to get this work I have to patch optimizer where it calls get_key_image() and see if key is HA_UNIQUE_HASH. If yes then instead of get_key_image just use memcpy(key, field->ptr(), field->pack_length()); this wont copy the actual data, but we do not need actual data. I will patch handler methods like ha_index_read, ha_index_idx_read, multi_range_read_info_const basically handler methods which are related to index or range search. In these methods i need to calculate hash, which I can calculate from key_ptr but key_ptr doe not have actual data(in case of blobs etc).So to get the data for hash, I will make a field clone of (a,b,c etc) but there ptr will point in key_ptr. Then field->val_str() method will work simply and i can calculate hash. And also I can compare returned result with actual key in handler method itself. What do you think of this approach ?
Looks simpler, agree. The length of the keypart should not matter, because it should never be used. May be it would be good to set it to -1 as it might help to catch errors (where it is erroneously used). I didn't understand why you need to clone fields though :( Regards, Sergei Chief Architect MariaDB and security@mariadb.org