What are distributed master-less file storages?

Question

Distributed Computing

Good day to all!
Does anyone know of a highly available distributed file storage?
Specificity:

• master-less, names-less - only one type of nodes. No separate metadata nodes etc. (good example: cassandra or cockroachdb)
  
• record source - always only one. No concurrent conflicting read/write operations
  
• number of nodes from 2 to 100
  
• nodes are weakly powerful, can appear and disappear (not instantly - you can count on synchronization, but the configuration is not fixed)
  
• write and read latency is not critical within reason (i.e. not HFT, but not batch mode either)
  
• redundancy support, or rather, data recovery guarantees
  
• linux only - windows and mac support is not important
  

Optional:

• sharding - i.e. linear increase in available size depending on the number of connected nodes
  
• file api, but own utilities are also suitable
  

Explored:

• gluster with replica - does not allow nodes to be added more than replica proof
  
• xtreemefs - requires separate nodes for proof metadata
  
• cockroachdb* - requires at least 3 nodes for raft proof
  
• cassandra* - java, quite demanding on resources
  

* it is clear that writing a wrapper on the database yourself and / or making a footcloth from scripts for synchronization is a matter of 2-3 days, but it will be a bicycle.

Answer 1

You can look at https://minio.io/ a la your private S3, there is no master, it works for reading if half of the nodes are lost and for writing if one less than half is lost, it can heal. Maybe at least one node, but fault-tolerance algorithms start working from 4 nodes. Perhaps this is not at all what you had in mind.

Answer 2

read in the direction of ceph, if something is not large, you can drive at 1gb / s, if you need HL directly, then you can’t do without 10gb / s

Answer 3

https://habr.com/company/yandex/blog/228389/
Elliptics from Yandex. How to use it to create your own fault-tolerant storage