I think the confusion regarding the reduced latch contention (if I understood it correctly :P) in case of multiple buffer pools may stem from the fact that this lock pic being shown on the page table in the slide gives inadvertently the impression that we are latching a particular frame when, in fact, it is the entire data structure. Is that correct? BTW, insanely good lectures Mr. Pavlo. Even you probably cannot fathom the full extent of impact making these lectures available for free will have over time. I just wanna say: Thank you!!!
52:16 Why if the data were already in the buffer the execution time of the query is still pretty much the same as before when Postgres did the complete scanning.
1:01:54 Isn't the main problem with LRU compared to the Clock algorithm, is that you need to maintain the order of the pages, so you actually know what *is* the least recently used page at any given moment?
Interesting point which however I think may not be the real issue. Clock also needs an ordering to sweep over all frames before starting the second trip. Hence ordering is needed there too. The page table can be dictionary to keep track of pages in memory and pages themselves can be a linked list for this kinda thing. For LRU you'd just pop off the head in constant time and for clock you'd just drop all the nodes that were set at zero each of which is also a constant time operation. Hence, i dont see why LRU would be more trouble than clock in this case. Or am I missing something?
You might be right. I thought we could use a fixed length array for Clock, so it wouldn't need to maintain a linked list structure. With LRU we have some additional pointer chasing, but probably it doesn't matter that much.
25:02 "SQL Server has the best query optimizer. It's phenomenal".
Watch this: th-cam.com/video/pQe1LQJiXN0/w-d-xo.html
I think the confusion regarding the reduced latch contention (if I understood it correctly :P) in case of multiple buffer pools may stem from the fact that this lock pic being shown on the page table in the slide gives inadvertently the impression that we are latching a particular frame when, in fact, it is the entire data structure. Is that correct?
BTW, insanely good lectures Mr. Pavlo. Even you probably cannot fathom the full extent of impact making these lectures available for free will have over time. I just wanna say: Thank you!!!
52:16 Why if the data were already in the buffer the execution time of the query is still pretty much the same as before when Postgres did the complete scanning.
Excellent Video!
1:01:54 Isn't the main problem with LRU compared to the Clock algorithm, is that you need to maintain the order of the pages, so you actually know what *is* the least recently used page at any given moment?
Interesting point which however I think may not be the real issue. Clock also needs an ordering to sweep over all frames before starting the second trip. Hence ordering is needed there too.
The page table can be dictionary to keep track of pages in memory and pages themselves can be a linked list for this kinda thing. For LRU you'd just pop off the head in constant time and for clock you'd just drop all the nodes that were set at zero each of which is also a constant time operation. Hence, i dont see why LRU would be more trouble than clock in this case.
Or am I missing something?
You might be right. I thought we could use a fixed length array for Clock, so it wouldn't need to maintain a linked list structure. With LRU we have some additional pointer chasing, but probably it doesn't matter that much.
FYI: the notes link is failing. Also: AWESOME lectures!
stop complaining
party samples/generators/etc. you use are.
What you play nice tutoet league on?
yo are you sober yet?