Distributed Snapshots: Determining Global States of Distributed Systems

Setup: 分布式snapshot算法，使用：apache flink, apache spark (structured streaming), ray

Reference review:

• https://matt33.com/2019/10/27/paper-chandy-lamport/

• Some examples that worth visit again: https://www.cs.princeton.edu/courses/archive/fall16/cos418/docs/P8-chandy-lamport.pdf

Main idea

• Problem: detecting global states

  • Which in terms help solve: "stable property detection" problem, e.g. "computation has terminated", "system is deadlocked", "all tokens in a token ring has disappeared"

  • Can also be used for checkpointing

• Propose an algorithm by which a process in a distributed system determines a global state of the system during a computation

Problem setup

• A process can record its own state and the messages it sends and receives, nothing else

  • No share clocks or memory

• To determine a global system state, a process p must enlist the cooperation of other processes that must record their own local states and send the recorded local states to p

• Problem: algorithms by which processes record their own states and the states of communication channels so that the set of processes and channel states recorded from a global system state

  • MUST run concurrently with, but not alter the underlying computation

    • Don't stop sending the messages, don't stop the application

• Example: taking photographies of migrating birds

• Define a class of problem

  • $y(S)= true / false$ for a global state S of D (distributed system)

    • y is a stable property of D if $y(S)$ implies $y(S')$for all global states $S'$ of D reachable from global state $S$ of D

    • i.e. if y is a stable property and y is true at a point in a computation of D, then y is true at all later points in that computation

  

  A stable property
• The problem of initiating the next phase of computation is not considered here

Model of a distributed system

• Labeled, directed graphs

  • Vertices: processes

    • Defined by a set of states, an initial state, and a set of events

    • An event $e$ in a process $p$ is an atomic action that may change the state of $p$ itself and the state of at most one channel $c$ incident on $p$

      • Event $e$ is defined by

        • 1) The process p in which the event occurs

        • 2) The state s of p immediately before the event

        • 3) The state s' of p immediately after the evetn

        • 4) The channel c (if any) whose state is altered by the event

        • 5) The message M, if any, sent along c (if c is a channel directed away from p) or received along c (if c is directed towards p)

      • M and c are a special symbol, $null$, if the occurrence of e does not change the state of any channel
      • Event can occur in global state $S$ if and only if

        • 1) The state of process p in global state S is s

        • and 2) if c is a channel directed towards p, then the state of c in global state S is a sequence of messages with M at its head

  • Edges: channels

    • State = sequence of messages sent along the channel, excluding the messages received along the channel

• Assumption

  • Channel: infinite buffers, error-free, delivery messages in the order sent

    
• A global state is a set of component process and channel states

  • Initial global state: state of each process is its initial state and the state of each channel is the empty sequence

  • $next(S, e)$= global state immediately after the occurrence of event e in global state S
  • = a sequence of events in component processes

    • computation of the system iff $e_i$ can occur in global state $S$

      • 
• Example 1

  • 
  • State transition illustration (four states: in-p, in-c, in-q, in-c')

    

Algorithm

• Motivation, outline, termination

Motivation

• How do global state recording algorithms work?

  • Each process: record its state

  • Two processes that a channel is incident on cooperate in recording the channel state

  • Require the recorded process and channel states form a "meaningful" global system state

• Example of transition

  • One scenario: inconsistency arises because the state of p is recorded before p sent a message along c and the state of c is recorded after p sent the message

    • n = number of msgs sent along c before p's state is recorded

      • 在 p 的状态记录之前，p 发往channel c的msg数

    • n' = number of msgs sent along c before c's state is recorded

      • 在 c 的状态记录之前，p 发往channel c的msg数

    • the recorded global state may be inconsistent if n < n'

      • 两个token的情况: p的状态记录在in-p中，其他的状态记录在in-c中

      • n = 0, n' = 1?

  • Another scenario: state of c is recorded in global state in-p, the system then transits to global state in-c, and the states of c', p, and q are recorded in global state in-c

    • The recorded global state shows no token in the system

    • Example suggests that the recorded global state may be inconsistent if the state of c is recorded before p sends a message along c and the state of p is recorded after p sends a message along c, that is, n > n'

  • Thus, a consistent global state requires (1) $n=n'$
• Definition

  • Let m be the number of messages received along c before q's state is recorded.

  • Let m' be the number of messages received along c before c's state is recorded

  • (2) $m=m'$: 与之前的分析相同
  • (3) $n'\geq m'$

    • I.e. # of msgs received along a channel cannot exceed the # of msgs sent along that channel in every state

  • (4) $n\ge m$
  • 理解

    • Channel要记录的state = list of msgs sent along the channel before the sender's state is recorded - list of msgs received along the channel before the receiver's state is recorded

      • 需要example理解这个内容

• (1) - (4) suggests a simple algorithm by which q can record the state of channel c

  • Process p sends a special message, called a marker, after the nth message it sends along c

  • The state of c is the sequence of messages received by q after q records its own state and before q receives the marker along c

  • To ensure (4), q must record its state, if it has not done so already, after receiving a marker along c and before q receives further messages along c

Algorithm

• Initiation

  • can be done by one or more processes, each of which records its state spontaneously, without receiving markers from other processes

• Marker-sending rule for a process p: for each channel c, incident on, and directed away from p

  • p sends one marker along c after p records its state and before p sends further messages along c

  • Or

    • p records its state and prepares a special marker message (differ from the application message)

    • p sends the marker message to all other processes (using N-1 outbound channels), in this case, only q

    • Start recording all incoming messages from channels $C_{ji}$for j not equal to i
  • Marker-receiving rule for a process q: on receiving a marker along a channel c

    

    • If not receive the marker message for the first time

      • Add all messages from inbound channels since we began recording to their states

• Termination of the algorithm

  • Need to ensure that

    • (L1) no marker remains forever in an incident input channel

    • (L2) it records its states within finite time of initiation of the algorithm

  • Rules

    • All processes have received a marker (and recorded their own state)

    • All processes have received a marker on the N - 1 incoming channels (and recorded their states)

    • Later, a central server can gather the partial state to build a global snapshot

Proof and properties ref. paper

• Casual consistency

  • Related to the Lamport clock partial ordering

  • An event is presnapshot if it occurs before the local snapshot on a process

  • Postsnapshot if afterwards

  • If event A happens casually before event B, and B is pre-snapshot, then A is too

Some questions

• Also assume no failures on the channels? All messages arrive, no duplicate, in order and such?

  • Too strong assumptions could reduce the deployability of the algorithm?

• What if the markers are lost? Retransmission? In-correct ordering?

• What if the topology is dynamically changing

• How fast does it take to form a global state? (with all the mechanisms of snapshotting, collecting, and putting together states from different components)

• Centralized controller seem to be more easy to implement and reason about