有懂bigtable ,hbase，c*的么？问一个timestamp的问题 - 未名空间MITBBS历史存档

国际科技财经博客移民网络热点娱乐民生时事公众号

Redian新闻

>未名空间

>Programming - 葵花宝典

有懂bigtable ,hbase，c*的么？问一个timestamp的问题

有懂bigtable ,hbase，c*的么？问一个timestamp的问题# Programming - 葵花宝典

t*r2015-03-23 07:03

1 楼

有懂bigtable ,hbase，c*的么？问一个timestamp的问题
bigtable里面timestamp的用法在hbase,c*里有么？
比如:by specifiying timestamp, I can get value of latest version that is
earlier than a specified timestamp.

p*22015-03-23 07:03

2 楼

对于一个key吧？C＊应该可以。

w*z2015-03-23 07:03

3 楼

C*每个column 有个timestamp, read return 的时候，如果cl >1,会比较timestamp of
the column from different node, in case of inconsistency, return the column
with the latest timestamp. client doesn't need to deal with the vector clock
crap. the last write wins.

【在 t**r 的大作中提到】

: 有懂bigtable ,hbase，c*的么？问一个timestamp的问题
: bigtable里面timestamp的用法在hbase,c*里有么？
: 比如:by specifiying timestamp, I can get value of latest version that is
: earlier than a specified timestamp.

g*g2015-03-23 07:03

4 楼

You want to have the timestamp as column name for indexing purpose.

of
column
clock

【在 w**z 的大作中提到】

: C*每个column 有个timestamp, read return 的时候，如果cl >1,会比较timestamp of
: the column from different node, in case of inconsistency, return the column
: with the latest timestamp. client doesn't need to deal with the vector clock
: crap. the last write wins.

f*t2015-03-23 07:03

5 楼

Hbase里有，get时可以指定一个timestamp

【在 t**r 的大作中提到】

w*z2015-03-23 07:03

6 楼

not sure what he is asking. if he needs timeseries data, then use timeuuid
as part of composite column name as you said.

【在 g*****g 的大作中提到】

: You want to have the timestamp as column name for indexing purpose.
:
: of
: column
: clock

g*g2015-03-23 07:03

7 楼

He basically wants to do range query on timestamp.

timeuuid

【在 w**z 的大作中提到】

: not sure what he is asking. if he needs timeseries data, then use timeuuid
: as part of composite column name as you said.

B*g2015-03-23 07:03

8 楼

c*如果还没有被彻底删掉，可以用sstable2json自己慢慢找，嘿嘿

【在 t**r 的大作中提到】

f*42015-03-23 07:03

9 楼

timestamp怎么同步？就是ntpd也有几十～一百ms的延时。多DC下面怎么搞？
就算不把C*当queue用这return the column with the latest timestamp还是需要
timestamp是正确的才行

【在 g*****g 的大作中提到】

: You want to have the timestamp as column name for indexing purpose.
:
: of
: column
: clock

g*g2015-03-23 07:03

10 楼

Timestamp can be marked on your application cluster. If you are doing mult-
DC and you are sensitive to ms order difference, you are probably doing it
wrong to begin with. i.e. You shouldn't do multi-DC for stock exchange.

【在 f****4 的大作中提到】

: timestamp怎么同步？就是ntpd也有几十～一百ms的延时。多DC下面怎么搞？
: 就算不把C*当queue用这return the column with the latest timestamp还是需要
: timestamp是正确的才行

f*42015-03-23 07:03

11 楼

好吧，我没讲清楚：C* in case of inconsistency, return the column with the
latest timestamp
C*自己解决inconsistency就是假设每个record的timestamp是sync的，这不是我应用需
要到ms精确度。
App cluster来做timestamp。你在app cluster里面怎么做同步？

【在 g*****g 的大作中提到】

: Timestamp can be marked on your application cluster. If you are doing mult-
: DC and you are sensitive to ms order difference, you are probably doing it
: wrong to begin with. i.e. You shouldn't do multi-DC for stock exchange.

g*g2015-03-23 07:03

12 楼

app cluster 总是 hit一个 node, 自然不用同步。如果你在两个DC同时写同一个
column, 结果的确有不确定性。但通常你不在乎写的是哪个。

【在 f****4 的大作中提到】

: 好吧，我没讲清楚：C* in case of inconsistency, return the column with the
: latest timestamp
: C*自己解决inconsistency就是假设每个record的timestamp是sync的，这不是我应用需
: 要到ms精确度。
: App cluster来做timestamp。你在app cluster里面怎么做同步？

w*z2015-03-23 07:03

13 楼

In most of cases, C* generates timestamp at server side for you. Make sure
your NTP is setup correctly which is a prerequisite for C* cluster
It's not advisable to generate timestamp at client side and pass it in with
your request.

【在 f****4 的大作中提到】

g*g2015-03-23 07:03

14 楼

Hmm, looks like cql has better support for time uuid now. There's a time
server side time uuid is not available.

with

【在 w**z 的大作中提到】

: In most of cases, C* generates timestamp at server side for you. Make sure
: your NTP is setup correctly which is a prerequisite for C* cluster
: It's not advisable to generate timestamp at client side and pass it in with
: your request.

f*42015-03-23 07:03

15 楼

我的感觉就是如果你的应用对NTP那100ms左右的误差引起的不一致很敏感的话C*就不是
你正确的选择
C*解决不一致性的时候是没办法100%保证正确的。CAP的trade off。

with

【在 w**z 的大作中提到】

w*z2015-03-23 07:03

16 楼

Or you can try to go down the vector clock hell like Riak.

【在 f****4 的大作中提到】

: 我的感觉就是如果你的应用对NTP那100ms左右的误差引起的不一致很敏感的话C*就不是
: 你正确的选择
: C*解决不一致性的时候是没办法100%保证正确的。CAP的trade off。
:
: with