blob: 8c7d12fc43e8a38070753cc55d18bdf1bbcb003a (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
Should probably just UUID all the (public) ids.
Instead of having a separate id pointer table, could have an extra "mutable"
public ID column (unique, indexed) on entity rows. Backend would ensure the
right thing happens. Changelog tables (or special redirect/deletion tables)
would record changes and be "fallen through" to.
Instead of having merge redirects, could just point all identifiers to the same
revision (and update them all in the future). Don't need to recurse! Need to
keep this forever though, could scale badly if "aggregations" get merged.
Redirections of redirections should probably simply be disallowed.
"Deletion" is really just pointing to a special or null entity.
Trade-off: easy querying for common case (wanting "active" rows) vs. robust
handling of redirects (likely to be pretty common). Also, having UUID handling
across more than one table.
## Scaling database
Two scaling issues: size of database due to edits (likely billions of rows) and
desire to do complex queries/reports ("analytics"). The later is probably not a
concern, and could be handled by dumping and working on a cluster (or secondary
views, etc). So just a distraction? Simpler to have all rolled up.
Cockroach is postgres-like; might be able to use that for HA and scaling?
Bottlenecks are probably complex joins (mitigated by "interleave"?) and bulk
import performance (one-time?).
Using elastic for most (eg, non-logged-in) views could keep things fast.
Cockroach seems more resourced/polished than TiDB?
|
