Re: Horizontal Write Scaling

Поиск
Список
Период
Сортировка
От Eliot Gable
Тема Re: Horizontal Write Scaling
Дата
Msg-id AANLkTimGuETOHVntCxoBDyZk28JRTVTC5T1mz4mQkj54@mail.gmail.com
обсуждение исходный текст
Ответ на Re: Horizontal Write Scaling  (Robert Haas <robertmhaas@gmail.com>)
Список pgsql-hackers
Thanks, everyone, for all the feedback! I am nowhere near a database expert yet, but you guys have been very helpful in clearing up some of my confusion. I have checked out Postgres-XC and it looks like the version 1.0 that will be released soon probably covers everything I have been looking for in terms of Postgres capabilities. The big ones are write scaling, read scaling, consistent view of data between each server, and HA capabilities. Last time I looked at Postgres-XC was probably a year ago, and it was nowhere close to what I was looking for at the time, and I forgot all about it. Now, it looks like a real contender. 

I was aware of Postgres-R and was actually thinking I might be able to get away with using that, but the project I am working on does a substantial amount of writing, as well as being CPU intensive. Each query executes a stored procedure which is about 2,500 lines long and pulls data from about 80 tables to compute a final result set. That final result set is returned to the requester, and is also written into 3 tables (while still inside the original transaction). One of those tables gets one row while the other two get 6 - 15 rows per query. I execute hundreds of these queries per second. So, I need to be able to spread the load across multiple boxes due to CPU usage, but still have a consistent view of the written data. Using conventional drives, I would saturate the disk I/O pretty quickly on commodity hardware. With normal multi-master replication, the cost of making sure I have enough disk I/O on each server is way more than I have the budget for. With a write scaling solution, it suddenly looks affordable. I was looking at maybe getting a single shared RAID array with some enterprise-class SSDs that could guarantee writes even during a power failure. I was hoping I could find something that would let multiple Postgres instances share that disk array as it would be more cost effective to get both the CPU power and Disk I/O I needed than sticking such a RAID array in each and every server I was going to spread load across. Postgres-XC actually makes it look even more affordable, as I now probably no longer need to consider SSDs, or at least I don't need to consider a RAID 10 array of 4 or more SSDs per box. I can probably do RAID 1 with 2 drives per box and have plenty of Disk I/O available for the amount of CPU power I would have in the boxes.

So, thanks again for the feedback.

В списке pgsql-hackers по дате отправления:

Предыдущее
От: Bruce Momjian
Дата:
Сообщение: Re: changing MyDatabaseId
Следующее
От: Bruce Momjian
Дата:
Сообщение: Re: Amazon now supporting GPU focused EC2 instances