C[omp]ute

From: "andrew cooke" <andrew@...>

Date: Fri, 7 Apr 2006 11:37:11 -0400 (CLT)

Listened to a talk on SDSS - http://cas.sdss.org/dr4/en/ - at work
yesterday.  These are my notes:

- 35 queries
The MS guru database chap asked them for 20 queries before designing the
database.  That grew to 35 that they now repeatedly run as benchmarks
after upgrades.  I was surprised at the number (20 seems a lot, and it got
bigger).  Good way to get non-experts to talk about the data model.

- keep all versions (inc bugs)
They had a fixed set of data they wanted to put on the web.  They did
that, and then started finding ways to improve things.  Great, but old
versions must stay - people are using the data in long term projects.

- raw sql
- user tables
Got burnt very early with OODB.  They do everything in SQL.  Data
remediation.  User's have their own scratch tables.  This is a big point
of conflict with opinions in our team.

- two phase loader - chunked
- first stage no indices
- verification in sql
- parallel loading
- second stage faster once data trusted
First loading stage takes raw data and builds index-free tables.  Data are
then remediated.  Second stage moves remediated data into indexed tables.

- sql nice for //n (cpus and disks)
- as many volumes as processors
- single table scans are slowest - disk read limited
Avoid big disks.  Parallelize across disks and processors.

- 2Mb crossover (file v sql)
That's pretty big for a blob, but much less than our binary data (images).

- submission queues to channel user expectations (slow web pages bad)
Batch processing is batch processing.  Admit it and make it clear to your
users.

- spatial features surprisingly popular
- "ferris wheel" scan - sequence of filters; scan database again and again
(for cross-matching)
Lots of details about spatial indexing and searching.  2D indexing is
hard.  If you often end up doing a scan, optimize for scans.  Ferris whell
model is repeated scanning, in chunks.  Scan a bunch of queries together. 
Queue queries for a chunk; go through the datbase scanning one chunk at a
time.

Andrew