Why TokuDB doesn't need double write?

[郭枫]

Since TokuDB doesn't need double write in InnoDB, how TokuDB gurantee atomic write ?（NVMFS and ZFS support atomic write so in these cases InnoDB doesn't need double write ）

[Leif Walsh]

Tokudb's block allocation is copy-on-write, so there is always a valid version of a block somewhere. Tokudb won't benefit from fs or hardware atomic writes, and anyway their block sizes might be too small. In any case they're probably just logging or doing copy on write themselves, so it's probably better to turn off those features if you're using tokudb. 

--
Cheers,
Leif

On Mon, Aug 24, 2015 at 10:09 PM, 郭枫 <leafo...@gmail.com> wrote:

Since TokuDB doesn't need double write in InnoDB, how TokuDB gurantee atomic write ?（NVMFS and ZFS support atomic write so in these cases InnoDB doesn't need double write ）

--
You received this message because you are subscribed to the Google Groups "tokudb-user" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tokudb-user...@googlegroups.com.
To post to this group, send email to tokud...@googlegroups.com.
Visit this group at http://groups.google.com/group/tokudb-user.
For more options, visit https://groups.google.com/d/optout.

[郭枫]

Do you mean?

For dml ,new block is copied and modified, even server crash, just new block is partial write, old block is still  intergrity, and redo log could be applied to old block to recovery?

在 2015年8月25日星期二 UTC+8上午10:32:17，Leif Walsh写道：

[Leif Walsh]

Exactly

--
Cheers,
Leif

[郭枫]

Why TokuDB use COW other than InnoDB's double write?It seems IO overhead is same:In InnoDB,there are two IO(write);In TokuDB there are also two IO ( one read for old data+one write for new data)~I think reason is COW support snapshot,so MVCC is simple to implement,is that so?

[Leif Walsh]

Innodb's pages are fixed-length so double-write makes more sense. Tokudb pages may change size depending how they compress, so COW is more suitable (and simpler to get right, IMO). 

--
Cheers,
Leif

[Leif Walsh]

Also, for innodb, it may cost a read plus two writes...

--
Cheers,
Leif

On Thu, Aug 27, 2015 at 10:46 PM, 郭枫 <leafo...@gmail.com> wrote:

[郭枫]

Thx~

Since variable page size is useful and flexible,do you know why InnoDB's page size is fixed in 4k、8k、16k、32k、64k........?

[Leif Walsh]

Fixed-size allocation is very simple. Variable-size allocation is complex and full of tradeoffs. 

--
Cheers,
Leif

[郭枫]

Thx~