@amitpr@HeyGarrison I think this is confusing two subtly different things. Some providers equate a vCPU to a hyperthread, but vCPUs are a completely different abstraction in the hypervisor/virtualization stack. A vCPU in KVM is basically the execution state of an abstract CPU core
@samlambert@BenjDicken@hazzngn That PG version doesn’t even have numa support and they’re not running on the largest azure instance. Can probably scale 2-3x further with dual xeon 6s
@hot_labor@rocketdrifterr@MartinShkreli Ppl were doing this a decade ago when jvm implementations weren’t great. Look up “Mechanical Sympathy java”. Today, you can easily run low latency — sub 1ms — apps with zgc.
@piyush_says@FranckPachot@DominikTornow No, the implication of my point is that once you have parallel writes there is no guarantee on order of events. For example, a remote write can complete faster than a local log write. Therefore, the replica that has the longest log prefix isn’t necessarily the leader.
@definekuujo Variants and optimizations. I’m not really familiar with raft, but have a lot of experience with Paxos. To really internalize and build intuition you really need to write inductive proofs for specs and debug/maintain real systems.
@definekuujo Distributed consensus is extremely hard to implement correctly and Lamport’s first two papers don’t help being really abstract. Arguably, he didn’t describe a protocol, but rather common properties that a family of protocols should have. This leaves a lot of space for many
@FranckPachot@DominikTornow This is incorrect. It depends on whether the commit optimization is being used or not. For example, the raft paper says that the leader first needs to apply the log entry locally before sending it to the quorum, but in the phd thesis it mentions a parallel quorum write
@chrisprucha@wagieeacc You thought he was having PTSD from prison, but you obviously haven’t seen the hater hotline videos pre-prison, or the mark cuban/jason saga. The man is not gonna take shit, he’s a straight up G
@jbensnyder@BillSaiWen@terjeholden It’s most certainly not DRAM. Even with +50% discount on DDR4 ECC the cost of ram alone would probably be more than 300mil, which is what they are reporting for the whole cluster.
@MarkCallaghanDB That doesn’t mean much because Glibc malloc’s defaults are not that great, you need to tune MALLOC_TRIM_THRESHOLD_ and MALLOC_ARENA_MAX . Sure Jemalloc/tcmalloc/minalloc can have better performance, but equating it to OOM is not a serious point