Content
@
0 reply
0 recast
0 reaction
Shane da Silva
@sds
I'm behind the times—apparently folks have been running Postgres on unikernels for quite some time†. https://www.prisma.io/blog/announcing-prisma-postgres-early-access https://nanovms.com/dev/tutorials/running-postgres-as-a-unikernel †
2 replies
3 recasts
23 reactions
Ryan
@ryansmith
is the idea that if you have file systems that are horizontally scalable and provide low latency, infinite IO then you would want low latency compute to spin up postgres to process queries on demand?
1 reply
0 recast
1 reaction
Shane da Silva
@sds
That's not what I immediately took away, but yes that seems like a potential advantage. In my mind the key point was that the benefits of unikernels, namely: - Reduced resource footprint and improved perf (no normal OS overhead of daemons, no kernel/user-space separation and thus less overhead from context switching or syscalls) - Fast boot and shutdown times (not a full OS) - Direct hardware access ...means that we can more densely pack a single baremetal machine with potentially thousands of database instances, since computation can be performed with a greater degree of liquidity (resources for different instances are only used when they are actually needed).
2 replies
0 recast
2 reactions
Mikko
@moo
I remember working on Oracle on Linux in 00s and it had direct block device access, no FS. When you exclusive lock a block device you can skip a bunch of syscalls?
1 reply
0 recast
2 reactions
Ryan
@ryansmith
Might also be able to do better caching and readahead
0 reply
0 recast
2 reactions
