veyepar: Vaultaire: a data vault for system metrics, backed onto Ceph

Author(s):	Andrew Cowie
Location	OGGB3
Date	jan Thu 15	Days Raw Files
Start	15:40	First Raw Start	15:09
Duration	0:45:00	Offset	0:30:07
End	16:25	Last Raw End	16:35
Chapters	00:00
Total cuts_time	50 min.

http://lca2015.linux.org.au/schedule/30074/view_talk
raw-playlist raw-mp4-playlist encoded-files-playlist host public tweet mp4 svg png

assets release.pdf Vaultaire_a_data_vault_for_system_metrics_backed_onto_Ceph.json
logs
Admin: episode episode list cut list raw files day marks day marks day image_files

State:

Locked:

clear this to unlock

Locked by:

user/process that locked.

Start:

initially scheduled time from master, adjusted to match reality

Duration:

length in hh:mm:ss

Name:

Video Title (shows in video search results)

Emails:

email(s) of the presenter(s)

Released:

has someone authorised pubication

Normalise:

Channelcopy:

m=mono, 01=copy left to right, 10=right to left, 00=ignore.

Thumbnail:

filename.png

Description:

markdown

Before you can do serious analytics you need to have serious data. You can't go averaging it away just to save space; you need the raw data.

Vaultaire is a distributed metrics store. It is fault tolerant, performs well, is space efficient, and — since it stores data directly in Ceph — scales horizontally as your cluster grows.

Most systems crumble when faced with the thousands of metrics per second typically being gathered in a production environment. Ceph is good at absorbing lots of writes, and is especially good at handling randomly distributed reads.

You may have experienced Ceph via it's S3-compatible HTTP gateway or via the RBD block device service. You may not know these are built atop a simple library that handles talking to the storage cluster and locating your data, librados, and that's how Vaultaire stores its time-series data.

Using librados directly opens some pretty cool possibilities. It's powerful, but low-level, and as you'd expect there are gotchas. We'll talk about building a librados-based application, lessons learned relying on Ceph in production, and the challenges of building predictive modelling tools on top of a metrics store that doesn't throw data away.

Comment:

production notes