michaelsevilla / malacology-popper

2 stars 0 forks source link

Add Performance Discussion #21

Closed michaelsevilla closed 8 years ago

michaelsevilla commented 8 years ago

We need to show that object classes solve a specific problem.

Use case: implement zlog with a proxy

  1. serialize over the metdata server
  2. interpose a new layer (like proxy) to do pipelining
    • the only way to do that in Ceph is with an IO
    • even if we do the same optimizations (e..g, batch up the IOs), what kind of performance hit do we get?
    • we would see that you need an entire networking and transactions stack; you would have to use the object interface on the OSD level -- you can't use any of the object classes
    • is this even possible? Best case scenario: prove that it is impossible

Reference Neha's HDFS HA namenode example