1 |
On 28/05/2014 13:42, Joost Roeleveld wrote: |
2 |
>>> Currently, I do the following: |
3 |
>>> > > Every year, a full backup |
4 |
>>> > > Then, every month, I have an incremental based on either the yearly or |
5 |
>>> > > previous monthly. |
6 |
>>> > > Ditto for the weekly (but then based on monthly or weekly) |
7 |
>>> > > And again for the daily. |
8 |
>> > |
9 |
>> > OK, that makes sense. |
10 |
>> > |
11 |
>> > It reminds me of an issue my wife had with the data warehouse when she |
12 |
>> > worked at the bank. In a nutshell, they needed backups but backups were |
13 |
>> > impossible to achieve because physics says so. They needed to get data |
14 |
>> > off the disk 4 times faster than data comes off a disk - SCSI limits |
15 |
>> > being rather hard limits :-) That opinion didn't go down well when I |
16 |
>> > offered it |
17 |
> Haha :) |
18 |
> I know the feeling. |
19 |
> I'd love to know the final solution they came up with. |
20 |
|
21 |
It was clever magic with LVM snapshots, but I'm not privy to the details |
22 |
- it was a bank after all and there's only so much Mrs Alan and the |
23 |
sysadmins could tell me. |
24 |
|
25 |
But it went something like this: |
26 |
|
27 |
Take a snapshot and copy that data to the SAN. This takes days or weeks |
28 |
and it's only ever done once. Thereafter, snapshots only. The backup |
29 |
system would take that last full + incrementals and create a new full |
30 |
for it's own use, this process runs independent from everything else and |
31 |
must take as long as it takes while the db carries on doing it's thing. |
32 |
If two backup jobs start to overlap in time, one of them gets discarded. |
33 |
|
34 |
Quite a clever scheme actually and it relies on storage being shared on |
35 |
the SAN to work, plus some in-house magic to get the backup system to |
36 |
recognise and use LVM snapshots natively. I believe performance impact |
37 |
was kept to acceptable levels by cleverly setting priorities - |
38 |
read/writes by the db take priority over backup reads which has to take |
39 |
a back seat and just wait it's turn. |
40 |
|
41 |
|
42 |
> |
43 |
>> > The solution was to do it much like your plan above. |
44 |
>> > With the benefit that the infrequent full backups would be done on a |
45 |
>> > fixed schedule in a change window with X hours downtime that was known |
46 |
>> > well in advance. |
47 |
> Using snapshots, the downtime is the same couple of minutes each night. |
48 |
> The problem is that during the backup, the performance of the server is |
49 |
> impacted. For a full backup, that means weeks... |
50 |
|
51 |
|
52 |
-- |
53 |
Alan McKinnon |
54 |
alan.mckinnon@×××××.com |