You're right on the etcd bits, but I think that EndpointSlices is also to reduce the size of what is sent over the network, see https://github.com/kubernetes/enhancements/blob/master/keps/sig-network/20190603-EndpointSlice-API.md#single-endpoint-update … . That said, I agree that it's a hack but probably too late to find different solutions 
-
-
-
The question is why it dents so much data. Endpoints should be a list of ip+port and not the full bloat object. How much data can this be? (32+16)*N Bit, N is number of nodes + fixed size header, which means 10k nodes is less than 60kB
- Još 3 druga odgovora
Novi razgovor -
-
-
Is there an efficient way to implement watches in Postgres (that is, without polling)? That seems to be the major reason for choosing etcd (perhaps also HA, but I'm less familiar about what Postgres can do here).
-
Watch can be done by trigger. For HA, run patroni. The best is that join could be used to do efficient label selector queries, which are the most inefficient query in etcd (think how to create endpoints from service and pods). EndpointSlices are a work around for this.
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
he is right, hacks like endpointSlices wouldn’t have to be done, if we were using Postgres for example