Try   HackMD

Introduction

This repository demonstrates how to use Wetware to write distributed systems. It contains a simple application that schedules jobs for execution in a distributed worker pool. A single scheduler node exposes a REST API for scheduling "tasks", and distributes them pseudorandomly to workers via Cap'n Proto RPC. The scheduler discovers worker nodes using Wetware's zero-conf clustering API, and can adapt to works joining and leaving the network dynamically.

Running it locally

Installation

  1. Clone the repository: 
    ​​​git clone https://github.com/evan-schott/ww-scheduler.git
​​​

First run

  1. Build the single application binary: 
    ​​​cd ww-scheduler && go build -o ww-sched cmd/scheduler/main.go
  2. Start the gateway node: 
    ​​​./ww-sched --gateway
​​​#Example Output: 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t
  3. Start a worker and add it to the gateway's pool 
    ​​​./ww-sched --dial 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t

The application is now deployed and ready to service requests.

Under the hood, Wetware has automatically provided the following:

  • Peer discovery: Wetware peers automatically find each other via a configurable bootstrap system. By default, peers locate each other using a protocol that is similar to mDNS, but more scalable.
  • Clustering: Wetware peers automatically connect to form an efficient overlay network, which they use to maintain mutual awareness. Each peer is able to detect the presence of other peers, and negotiate network connections.
  • Routing: Peers can route data and network connections to each other, using each peer's unique identifier.
  • Services: Wetware provides a set of cluster-oriented services upon which to build applications. The present demo makes use of Wetware's native support for Cap'n Proto's RPC, which provides a compact data encoding, a latency-aware RPC API, and security through object-capabilities.

Architecture

Architecture of the Wetware Stack

  • Bootstrapping
    • Explain the bootstrapping problem
      • Peers can use p2p algos to find other peers
      • … but how do I discover the first peer?
    • About ww's pluggable bootstrap architecture
      • Ww's stance: no one-size-fits-all solution
      • Provide Go interface called Discovery; implementations include
        • Static bootstrap nodes
        • Multicast UDP
        • Public boot server (centralized)
        • Crawl IP range (CIDR block)
        • List of static boot nodes
        • User-defined implementations
    • About the multicast udp boot service
      • How it works
      • Relation to mDNS (= "why not just use mDNS?")
      • Brief note about docker: how and why it works in containers
      • Brief note about UDP multicast in cloud environments:
        • It's generally not enabled
        • Use another bootstrap service, e.g.:
          • public boot server (centralized)
          • list of static boot nodes
          • crawl a CIDR block (surprisingly good experience)
  • Clustering
    • Problem statement: how can I group a set of computers together in a way that allows each computer to enumerate its peers?
    • Ww's stance:
      • There is a one-size-fits all solution
      • … but it requires sacrifices
    • Solution is: PA/EL system
      • Reliable systems can be built on top of unreliable parts
      • Model the cluster as:
        • Unreliable
        • Asynchronous
        • Stateful
        • Concurrent
    • Sacrifices are:
      • Consistency is an application concern
      • Users have to know what they are doing
    • Rationale:
  • Routing
    • Problem statement: how can peers form connections with each other and transmit data?
    • Ww's stance: this is provided to us by libp2p
    • Solution is:
      • "Process addressing"
        • Each process has its own peer.ID
        • Each process multiplexes data into "streams"
        • Each "stream" has a handler, analogous to HTTP
      • DHT
        • Immutable "records"
          • High persistence: records persist after hosts leave
          • Low liveness: updates slow to propagate
        • Records are
          • Immutable
          • Cryptographically signed & authenticated
      • PubSub
        • High-liveness "stream" of raw data
          • High liveness: responsive/updates propagate quickly
          • Low persistence: hosts must be online
        • Stream is pubsub topic
          • Asynchronous
          • Unreliable
          • Cryptographically signed & authenticated
      • Routed Host
        • Network interface for managing multiplexed streams between peers
        • Employs DHT & PubSub to route by global peer ID
  • Services
    • Problem Statement: what global services can be provided to the application?
    • Ww's stance:
      • The real world is a PA/EL system
      • never lie to the user about that.
    • Solution is: API design
      • Data is immutable
      • RPC calls are asynchronous
      • Security is provided by object capabilities
    • Sacrifices are:
      • APIs cannot offer any consistency guarantees
    • Rationale:
      • Consistent APIs can be built from inconsistent APIs
      • We can offer a standard library of commonly-used consistency protocols
        • Two-Phase Commit
        • Three-Phase Commit
        • Synchronous Consistency
          • RAFT
          • Paxos
        • Non-Byzantine Eventual Consistency
          • CRDT Log
        • Byzantine Eventual Consistency
          • Block Chain (probably L2 integration of some kind)

Application Architecture

  • Gateway node:
    • Services used
      • Boot
      • Clustering
      • Routing
      • RPC
      • CSP
    • Application layer
      • HTTP REST server
      • Task Scheduling logic
  • Worker node
    • Services used
      • Boot
      • Clustering
      • Routing
      • RPC
      • CSP
    • Application layer
      • Cap'n Proto Worker capability
        • Executes WASM payload when told to by gateway node
      • WASM runtime
        • Executes arbitrary WASM bytecode

// In one shell to run gateway: (output will contain gateway peerID)
go run cmd/scheduler/main.go –gateway true

// In other shells to run workers
go run cmd/scheduler/main.go –dial [INSERT GATEWAY PEERID] 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t



# Running it with docker
TODO: add docker config instructions

# In more depth
## 1 Bootstrapping 
A node joins the wetware cluster by bootstrapping to a peer in the network. This procedure involves sending and listening for discover packets in the multicast group specified in the header flags. In order to view all the packets you download [casm](https://github.com/wetware/casm)and run `casm cluster ls` . 
<details><summary>Example Output</summary>

[0000] Got multicast packet (171 byte)
type: survey
namespace: ww
size: 171 bytes
peer: 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t
distance: 255

[0000] Got multicast packet (217 byte)
type: response
namespace: ww
size: 217 bytes
peer: 12D3KooWQdc1sHWeBqbWuCPZEG5gjLy4tWrc8ppKYTRAMq4upeTP

[0006] Got multicast packet (171 byte)
type: survey
namespace: ww
size: 171 bytes
peer: 12D3KooWQdc1sHWeBqbWuCPZEG5gjLy4tWrc8ppKYTRAMq4upeTP
distance: 255

[0006] Got multicast packet (217 byte)
type: response
namespace: ww
size: 217 bytes
peer: 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t

[0009] Got multicast packet (171 byte)
type: survey
namespace: ww
size: 171 bytes
peer: 12D3KooWKggVqaYBwJvCCFNV8zNqCQS6rbSMKTafjmMBH2pZpZ8t
distance: 255

[0009] Got multicast packet (217 byte)
type: response
namespace: ww
size: 217 bytes
peer: 12D3KooWQdc1sHWeBqbWuCPZEG5gjLy4tWrc8ppKYTRAMq4upeTP

​​
</details>

## 2. Capabilities
Capabilities are fundamental to the access control model of Wetware. A process must own a capability from another process in order to call RPCs. Capabilities can be broadcasted  as demonstrated with:
```go
// Create synchronous channel server on Gateway server and export it
// Now any node on network can send to our channel server by using the channel capability
ch := csp.NewChan(&csp.SyncChan{})
defer ch.Close(c.Context)
n.Vat.Export(chanCap, chanProvider{ch})

Capabilities can also be sent privately by way of transmission through a channel.

// Setup of the worker capability. Start a worker server, and derive a client from it.
server := worker.WorkerServer{}

// Block until we're able to send our worker capability to the
// gateway server. This is where the load-balancing happens.
// We are competing with other Send()s, and the gateway will
// pick one of the senders at random each time it handles an
// HTTP request.
e := capnp.Client(worker.Worker_ServerToClient(server))
err = ch.Send(context.Background(), csp.Client(e))

3. Channel Server

The channel server is a fundamental primitive for interprocess communication (IPC). Here are the primary differences between using PubSub and using channels for IPC:

Property Channel PubSub
Reliability Every object sent to the channel server will be held onto until it is received. Packets will be dropped once buffer capacity is hit.
Ordering If a receive request arrives at the channel server before another receive request, then it will return a value from a send request that arrived to the channel before the send request value returned to the later receive request No ordering guarantees.
Synchrony A call to send an object to the channel server will block until the value is successfully received. Likewise a call to receive an object from the channel will block until an object is available. No synchrony guarantees.
Medium Uses unicast Uses multicast

4. Layering (Libp2p, Cap'n Proto, Casm, Wetware)

Libp2p: (Networking stack)

  • "Host" abstraction to represent node in network.
  • "Connections" abstraction of transport layer link between two peers that can support multiple "Streams" on top of them to represent different communication topics.

Cap'n Proto: (RPC protocol)

  • "Capabilities" abstraction to represent permission for making calls on remote objects.
  • Architecture for configuring dynamic configuration of capabilities between hosts.
  • Efficient serialization scheme
  • Promise pipelining

Casm: (Low level cluster primitives)

  • Merger between Cap'n Proto and Libp2p layers to allow for capabilities sharing on cluster of Libp2p Hosts.
  • Bootstrap protocol for joining cluster to peers on cluster and forming cluster.

Wetware: (Cluster middleware)

  • Higher level primitives for IPC (PubSub and Channels), shared memory (Anchors), processes, and querying information about cluster state.

5. WASM

In this example, the "tasks" that can be scheduled on worker nodes take the form of hash puzzles. But really any .wasm file can be used as a task. The directions below outline how the tasks are compiled, and then how to make requests to the cluster to execute the tasks.

// To compile hash.go into a .wasm representation
tinygo build -o wasm/hash.wasm -target=wasi hash.go

// To schedule a difficulty 3 hash puzzle with seed "0", that will start 1 second after it is received, and repeat every 2 seconds for 3 iterations.
curl -X POST -H "Content-Type: multipart/form-data" \
     -F "id=task1" \
     -F "description=Hash puzzle" \
     -F "complete=0" \
     -F "duration=5" \
     -F "start=1" \
     -F "delay=2" \
     -F "repeats=3" \
     -F "wasm=@hash.wasm" \
     -F "input=0" \
     -F "difficulty=3" \
     http://localhost:8080/tasks

// To see the current state of jobs, and their progress towards completion
curl -X GET http://localhost:8080/tasks

In action

  • Link to video
    TODO: create video
Last changed by 
 
Louis Thibault
0
115

Read more

Wetware Cluster Membership

…

Mar 29, 2024
[ ]

[ ]

Nov 14, 2023
WASM-Based Precompiles for SUAVE

This memo is intended to summarize preliminary work on supporting WebAssmebly-based pre-compiled contracts in Flashbots’ SUAVE implementation. Two separate features have been undertaken:

Oct 4, 2023
Wetware

Decentralized Cloud architecture for Web3.

Sep 8, 2023
Read more from Louis Thibault
Published on HackMD