greg/machi
1
0
Fork 0
Code Issues 11 Pull requests 1 Projects Releases Packages Wiki Activity Actions
machi/doc/chain-self-management-sketch.org
Scott Lystig Fritchie 78f2ff4bbf Number section headings, clarify flapping behavior, add prototype notes 
Fix #+END_QUOTE typo
2015-03-14 12:06:50 +09:00

32 KiB
Raw Blame History

Machi Chain Self-Management Sketch

-- mode: org; --

1. Abstract

Yo, this is the first draft of a document that attempts to describe a proposed self-management algorithm for Machi's chain replication. Welcome! Sit back and enjoy the disjointed prose.

We attempt to describe first the self-management and self-reliance goals of the algorithm. Then we make a side trip to talk about write-once registers and how they're used by Machi, but we don't really fully explain exactly why write-once is so critical (why not general purpose registers?) … but they are indeed critical. Then we sketch the algorithm by providing detailed annotation of a flowchart, then let the flowchart speak for itself, because writing good prose is prose is damn hard, but flowcharts are very specific and concise.

Finally, we try to discuss the network partition simulator that the algorithm runs in and how the algorithm behaves in both symmetric and asymmetric network partition scenarios. The symmetric partition cases are all working well (surprising in a good way), and the asymmetric partition cases are working well (in a damn mystifying kind of way). It'd be really, really great to get more review of the algorithm and the simulator.

2. Copyright 

%% Copyright (c) 2015 Basho Technologies, Inc.  All Rights Reserved.
%%
%% This file is provided to you under the Apache License,
%% Version 2.0 (the "License"); you may not use this file
%% except in compliance with the License.  You may obtain
%% a copy of the License at
%%
%%   http://www.apache.org/licenses/LICENSE-2.0
%%
%% Unless required by applicable law or agreed to in writing,
%% software distributed under the License is distributed on an
%% "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
%% KIND, either express or implied.  See the License for the
%% specific language governing permissions and limitations
%% under the License.

3. Naming: possible ideas (TODO)

Humming consensus?

See On Consensus and Humming in the IETF, RFC 7282.

See also: On “Humming Consensus”, an allegory.

Foggy consensus?

CORFU-like consensus between mist-shrouded islands of network partitions

Rough consensus

This is my favorite, but it might be too close to handwavy/vagueness of English language, even with a precise definition and proof sketching?

Let the bikeshed continue!

I agree with Chris: there may already be a definition that's close enough to "rough consensus" to continue using that existing tag than to invent a new one. TODO: more research required

4. What does "self-management" mean in this context?

For the purposes of this document, chain replication self-management is the ability for the N nodes in an N-length chain replication chain to manage the state of the chain without requiring an external party to participate. Chain state includes:

  1. Preserve data integrity of all data stored within the chain. Data loss is not an option.
  2. Stably preserve knowledge of chain membership (i.e. all nodes in the chain, regardless of operational status). A systems administrators is expected to make "permanent" decisions about chain membership.
  3. Use passive and/or active techniques to track operational state/status, e.g., up, down, restarting, full data sync, partial data sync, etc.
  4. Choose the run-time replica ordering/state of the chain, based on current member status and past operational history. All chain state transitions must be done safely and without data loss or corruption.
  5. As a new node is added to the chain administratively or old node is restarted, add the node to the chain safely and perform any data synchronization/"repair" required to bring the node's data into full synchronization with the other nodes.

5. Goals

Better than state-of-the-art: Chain Replication self-management

We hope/believe that this new self-management algorithem can improve the current state-of-the-art by eliminating all external management entities. Current state-of-the-art for management of chain replication chains is discussed below, to provide historical context.

"Leveraging Sharding in the Design of Scalable Replication Protocols" by Abu-Libdeh, van Renesse, and Vigfusson.

Multiple chains are arranged in a ring (called a "band" in the paper). The responsibility for managing the chain at position N is delegated to chain N-1. As long as at least one chain is running, that is sufficient to start/bootstrap the next chain, and so on until all chains are running. (The paper then estimates mean-time-to-failure (MTTF) and suggests a "band of bands" topology to handle very large clusters while maintaining an MTTF that is as good or better than other management techniques.)

If the chain self-management method proposed for Machi does not succeed, this paper's technique is our best fallback recommendation.

An external management oracle, implemented by ZooKeeper

This is not a recommendation for Machi: we wish to avoid using ZooKeeper. However, many other open and closed source software products use ZooKeeper for exactly this kind of data replica management problem.

An external management oracle, implemented by Riak Ensemble

This is a much more palatable choice than option #2 above. We also wish to avoid an external dependency on something as big as Riak Ensemble. However, if it comes between choosing Riak Ensemble or choosing ZooKeeper, the choice feels quite clear: Riak Ensemble will win, unless there is some critical feature missing from Riak Ensemble. If such an unforseen missing feature is discovered, it would probably be preferable to add the feature to Riak Ensemble rather than to use ZooKeeper (and document it and provide product support for it and so on…).

Support both eventually consistent & strongly consistent modes of operation

Machi's first use case is for Riak CS, as an eventually consistent store for CS's "block" storage. Today, Riak KV is used for "block" storage. Riak KV is an AP-style key-value store; using Machi in an AP-style mode would match CS's current behavior from points of view of both code/execution and human administrator exectations.

Later, we wish the option of using CP support to replace other data store services that Riak KV provides today. (Scope and timing of such replacement TBD.)

We believe this algorithm allows a Machi cluster to fragment into arbitrary islands of network partition, all the way down to 100% of members running in complete network isolation from each other. Furthermore, it provides enough agreement to allow formerly-partitioned members to coordinate the reintegration & reconciliation of their data when partitions are healed.

Preserve data integrity of Chain Replicated data

While listed last in this section, preservation of data integrity is paramount to any chain state management technique for Machi.

Anti-goal: minimize churn

This algorithm's focus is data safety and not availability. If participants have differing notions of time, e.g., running on extremely fast or extremely slow hardware, then this algorithm will "churn" in different states where the chain's data would be effectively unavailable.

In practice, however, any series of network partition changes that case this algorithm to churn will cause other management techniques (such as an external "oracle") similar problems. [Proof by handwaving assertion.] See also: "time model" assumptions (below).

6. Assumptions

Introduction to assumptions, why they differ from other consensus algorithms

Given a long history of consensus algorithms (viewstamped replication, Paxos, Raft, et al.), why bother with a slightly different set of assumptions and a slightly different protocol?

The answer lies in one of our explicit goals: to have an option of running in an "eventually consistent" manner. We wish to be able to make progress, i.e., remain available in the CAP sense, even if we are partitioned down to a single isolated node. VR, Paxos, and Raft alone are not sufficient to coordinate service availability at such small scale.

The CORFU protocol is correct

This work relies tremendously on the correctness of the CORFU protocol, a cousin of the Paxos protocol. If the implementation of this self-management protocol breaks an assumption or prerequisite of CORFU, then we expect that the implementation will be flawed.

Communication model: Asyncronous message passing

Unreliable network: messages may be arbitrarily dropped and/or reordered

Network partitions may occur at any time
Network partitions may be asymmetric: msg A->B is ok but B->A fails

Messages may be corrupted in-transit

Assume that message MAC/checksums are sufficient to detect corruption
Receiver informs sender of message corruption
Sender may resend, if/when desired

System particpants may be buggy but not actively malicious/Byzantine

Time model: per-node clocks, loosely synchronized (e.g. NTP)

The protocol & algorithm presented here do not specify or require any timestamps, physical or logical. Any mention of time inside of data structures are for human/historic/diagnostic purposes only.

Having said that, some notion of physical time is suggested for purposes of efficiency. It's recommended that there be some "sleep time" between iterations of the algorithm: there is no need to "busy wait" by executing the algorithm as quickly as possible. See below, "sleep intervals between executions".

Failure detector model: weak, fallible, boolean

We assume that the failure detector that the algorithm uses is weak, it's fallible, and it informs the algorithm in boolean status updates/toggles as a node becomes available or not.

If the failure detector is fallible and tells us a mistaken status change, then the algorithm will "churn" the operational state of the chain, e.g. by removing the failed node from the chain or adding a (re)started node (that may not be alive) to the end of the chain. Such extra churn is regrettable and will cause periods of delay as the "rough consensus" (decribed below) decision is made. However, the churn cannot (we assert/believe) cause data loss.

The "wedge state", as described by the Machi RFC & CORFU

A chain member enters "wedge state" when it receives information that a newer projection (i.e., run-time chain state reconfiguration) is available. The new projection may be created by a system administrator or calculated by the self-management algorithm. Notification may arrive via the projection store API or via the file I/O API.

When in wedge state, the server/FLU will refuse all file write I/O API requests until the self-management algorithm has determined that "rough consensus" has been decided (see next bullet item). The server may also refuse file read I/O API requests, depending on its CP/AP operation mode.

See the Machi RFC for more detail of the wedge state and also the CORFU papers.

"Rough consensus": consensus built upon data that is visible now

CS literature uses the word "consensus" in the context of the problem description at http://en.wikipedia.org/wiki/Consensus_(computer_science)#Problem_description. This traditional definition differs from what is described in this document.

The phrase "rough consensus" will be used to describe consensus derived only from data that is visible/known at the current time. This implies that a network partition may be in effect and that not all chain members are reachable. The algorithm will calculate "rough consensus" despite not having input from all/majority/minority of chain members. "Rough consensus" may proceed to make a decision based on data from only a single participant, i.e., the local node alone.

When operating in AP mode, i.e., in eventual consistency mode, "rough consensus" could mean that an chain of length N could split into N independent chains of length 1. When a network partition heals, the rough consensus is sufficient to manage the chain so that each replica's data can be repaired/merged/reconciled safely. (Other features of the Machi system are designed to assist such repair safely.)

When operating in CP mode, i.e., in strong consistency mode, "rough consensus" would require additional supplements. For example, any chain that didn't have a minimum length of the quorum majority size of all members would be invalid and therefore would not move itself out of wedged state. In very general terms, this requirement for a quorum majority of surviving participants is also a requirement for Paxos, Raft, and ZAB.

(Aside: The Machi RFC also proposes using "witness" chain members to make service more available, e.g. quorum majority of "real" plus "witness" nodes and at least one member must be a "real" node. See the Machi RFC for more details.)

Heavy reliance on a key-value store that maps write-once registers

The projection store is implemented using "write-once registers" inside a key-value store: for every key in the store, the value must be either of:

  • The special 'unwritten' value
  • An application-specific binary blob that is immutable thereafter

7. The projection store, built with write-once registers

  • NOTE to the reader: The notion of "public" vs. "private" projection stores does not appear in the Machi RFC.

Each participating chain node has its own "projection store", which is a specialized key-value store. As a whole, a node's projection store is implemented using two different key-value stores:

  • A publicly-writable KV store of write-once registers
  • A privately-writable KV store of write-once registers

Both stores may be read by any cluster member.

The store's key is a positive integer; the integer represents the epoch number of the projection. The store's value is an opaque binary blob whose meaning is meaningful only to the store's clients.

See the Machi RFC for more detail on projections and epoch numbers.

The publicly-writable half of the projection store

The publicly-writable projection store is used to share information during the first half of the self-management algorithm. Any chain member may write a projection to this store.

The privately-writable half of the projection store

The privately-writable projection store is used to store the "rough consensus" result that has been calculated by the local node. Only the local server/FLU may write values into this store.

The private projection store serves multiple purposes, including:

  • remove/clear the local server from "wedge state"
  • act as the store of record for chain state transitions
  • communicate to remote nodes the past states and current operational state of the local node

8. Modification of CORFU-style epoch numbering and "wedge state" triggers

According to the CORFU research papers, if a server node N or client node C believes that epoch E is the latest epoch, then any information that N or C receives from any source that an epoch E+delta (where delta > 0) exists will push N into the "wedge" state and C into a mode of searching for the projection definition for the newest epoch.

In the algorithm sketch below, it should become clear that it's possible to have a race where two nodes may attempt to make proposals for a single epoch number. In the simplest case, assume a chain of nodes A & B. Assume that a symmetric network partition between A & B happens, and assume we're operating in AP/eventually consistent mode.

On A's network partitioned island, A can choose a UPI list of `[A]'. Similarly B can choose a UPI list of `[B]'. Both might choose the epoch for their proposal to be #42. Because each are separated by network partition, neither can realize the conflict. However, when the network partition heals, it can become obvious that there are conflicting values for epoch #42 … but if we use CORFU's protocol design, which identifies the epoch identifier as an integer only, then the integer 42 alone is not sufficient to discern the differences between the two projections.

The proposal modifies all use of CORFU's projection identifier to use the identifier below instead. (A later section of this document presents a detailed example.)

{epoch #, hash of the entire projection (minus hash field itself)}

9. Sketch of the self-management algorithm

Introduction

Refer to the diagram `chain-self-management-sketch.Diagram1.pdf`, a flowchart of the algorithm. The code is structured as a state machine where function executing for the flowchart's state is named by the approximate location of the state within the flowchart. The flowchart has three columns:

  1. Column A: Any reason to change?
  2. Column B: Do I act?
  3. Column C: How do I act?

States in each column are numbered in increasing order, top-to-bottom.

Flowchart notation

  • Author: a function that returns the author of a projection, i.e., the node name of the server that proposed the projection.
  • Rank: assigns a numeric score to a projection. Rank is based on the epoch number (higher wins), chain length (larger wins), number & state of any repairing members of the chain (larger wins), and node name of the author server (as a tie-breaking criteria).
  • E: the epoch number of a projection.
  • UPI: "Update Propagation Invariant". The UPI part of the projection is the ordered list of chain members where the UPI is preserved, i.e., all UPI list members have their data fully synchronized (except for updates in-process at the current instant in time).
  • Repairing: the ordered list of nodes that are in "repair mode", i.e., synchronizing their data with the UPI members of the chain.
  • Down: the list of chain members believed to be down, from the perspective of the author. This list may be constructed from information from the failure detector and/or by status of recent attempts to read/write to other nodes' public projection store(s).
  • P_current: local node's projection that is actively used. By definition, P_current is the latest projection (i.e. with largest epoch #) in the local node's private projection store.
  • P_newprop: the new projection proposal that is calculated locally, based on local failure detector info & other data (e.g., success/failure status when reading from/writing to remote nodes' projection stores).
  • P_latest: this is the highest-ranked projection with the largest single epoch # that has been read from all available public projection stores, including the local node's public store.
  • Unanimous: The P_latest projections are unanimous if they are effectively identical. Minor differences such as creation time may be ignored, but elements such as the UPI list must not be ignored. NOTE: "unanimous" has nothing to do with the number of projections compared, "unanimous" is not the same as a "quorum majority".
  • P_current -> P_latest transition safe?: A predicate function to check the sanity & safety of the transition from the local node's P_current to the P_newprop, which must be unanimous at state C100.
  • Stop state: one iteration of the self-management algorithm has finished on the local node. The local node may execute a new iteration at any time.

Column A: Any reason to change?

A10: Set retry counter to 0

A20: Create a new proposed projection based on the current projection

A30: Read copies of the latest/largest epoch # from all nodes

A40: Decide if the local proposal P_newprop is "better" than P_latest

Column B: Do I act?

B10: 1. Is the latest proposal unanimous for the largest epoch #?

B10: 2. Is the retry counter too big?

B10: 3. Is another node's proposal "ranked" equal or higher to mine?

Column C: How to act?

C1xx: Save latest proposal to local private store, unwedge, stop.

C2xx: Ping author of latest to try again, then wait, then repeat alg.

C3xx: My new proposal appears best: write @ all public stores, repeat alg

Flowchart notes

Algorithm execution rates / sleep intervals between executions

Due to the ranking algorithm's preference for author node names that are small (lexicographically), nodes with smaller node names should execute the algorithm more frequently than other nodes. The reason for this is to try to avoid churn: a proposal by a "big" node may propose a UPI list of L at epoch 10, and a few moments later a "small" node may propose the same UPI list L at epoch 11. In this case, there would be two chain state transitions: the epoch 11 projection would be ranked higher than epoch 10's projeciton. If the "small" node executed more frequently than the "big" node, then it's more likely that epoch 10 would be written by the "small" node, which would then cause the "big" node to stop at state A40 and avoid any externally-visible action.

Transition safety checking

In state C100, the transition from P_current -> P_latest is checked for safety and sanity. The conditions used for the check include:

  1. The Erlang data types of all record members are correct.
  2. UPI, down, & repairing lists contain no duplicates and are in fact mutually disjoint.
  3. The author node is not down (as far as we can tell).
  4. Any additions in P_latest in the UPI list must appear in the tail of the UPI list and were formerly in P_current's repairing list.
  5. No re-ordering of the UPI list members: P_latest's UPI list prefix must be exactly equal to P_current's UPI prefix, and any P_latest's UPI list suffix must in the same order as they appeared in P_current's repairing list.

The safety check may be performed pair-wise once or pair-wise across the entire history sequence of a server/FLU's private projection store.

A simple example race between two participants noting a 3rd's failure

Assume a chain of three nodes, A, B, and C. In a projection at epoch

  1. For all nodes, the P_current projection at epoch E is:

UPI=[A,B,C], Repairing=[], Down=[]

Now assume that C crashes during epoch E. The failure detector running locally at both A & B eventually notice C's death. The new information triggers a new iteration of the self-management algorithm. A calculates its P_newprop (call it P_newprop_a) and writes it to its own public projection store. Meanwhile, B does the same and wins the race to write P_newprop_b to its own public projection store.

At this instant in time, the public projection stores of each node looks something like this:

Epoch Node A Node B Node C
E UPI=[A,B,C] UPI=[A,B,C] UPI=[A,B,C]
Repairing=[] Repairing=[] Repairing=[]
Down=[] Down=[] Down=[]
Author=A Author=A Author=A
E+1 UPI=[A,B] UPI=[A,B] C is dead,
Repairing=[] Repairing=[] unwritten
Down=[C] Down=[C]
Author=A Author=B

If we use the CORFU-style projection naming convention, where a projection's name is exactly equal to the epoch number, then all participants cannot tell the difference between the projection at epoch E+1 authored by node A from the projection at epoch E+1 authored by node B: the names are the same, i.e., E+1.

Machi must extend the original CORFU protocols by changing the name of the projection. In Machi's case, the projection is named by this 2-tuple: 

{epoch #, hash of the entire projection (minus hash field itself)}

This name is used in all relevant APIs where the name is required to make a wedge state transition. In the case of the example & table above, all of the UPI & Repairing & Down lists are equal. However, A & B's unanimity is due to the symmetric nature of C's partition: C is dead. In the case of an asymmetric partition of C, it is indeed possible for A's version of epoch E+1's UPI list to be different from B's UPI list in the same epoch E+1.

A second example, building on the first example

Building on the first example, let's assume that A & B have reconciled their proposals for epoch E+2. Nodes A & B are running under a unanimous proposal at E+2.

E+2 UPI=[A,B] UPI=[A,B] C is dead,
Repairing=[] Repairing=[] unwritten
Down=[C] Down=[C]
Author=A Author=A

Now assume that C restarts. It was dead for a little while, and its code is slightly buggy. Node C decides to make a proposal without first consulting its failure detector: let's assume that C believes that only C is alive. Also, C knows that epoch E was the last epoch valid before it crashed, so it decides that it will write its new proposal at E+2. The result is a set of public projection stores that look like this:

E+2 UPI=[A,B] UPI=[A,B] UPI=[C]
Repairing=[] Repairing=[] Repairing=[]
Down=[C] Down=[C] Down=[A,B]
Author=A Author=A Author=C

Now we're in a pickle where a client C could read the latest projection from node C and get a different view of the world than if it had read the latest projection from nodes A or B.

If running in AP mode, this wouldn't be a big problem: a write to node C only (or a write to nodes A & B only) would be reconciled eventually. Also, eventually, one of the nodes would realize that C was no longer partitioned and would make a new proposal at epoch E+3.

If running in CP mode, then any client that attempted to use C's version of the E+2 projection would fail: the UPI list does not contain a quorum majority of nodes. (Other discussion of CP mode's use of quorum majority for UPI members is out of scope of this document. Also out of scope is the use of "witness servers" to augment the quorum majority UPI scheme.)

10. The Network Partition Simulator

Overview

The function machi_chain_manager1_test:convergence_demo_test() executes the following in a simulated network environment within a single Erlang VM:

Test the convergence behavior of the chain self-management algorithm for Machi.

  1. Set up 4 FLUs and chain manager pairs.
  2. Create a number of different network partition scenarios, where (simulated) partitions may be symmetric or asymmetric. (At the Seattle 2015 meet-up, I called this the "shaking the snow globe" phase, where asymmetric network partitions are simulated and are calculated at random differently for each simulated node. During this time, the simulated network is wildly unstable.)
  3. Then halt changing the partitions and keep the simulated network stable. The simulated may remain broken (i.e. at least one asymmetric partition remains in effect), but at least it's stable.
  4. Run a number of iterations of the algorithm in parallel by poking each of the manager processes on a random'ish basis to simulate the passage of time.
  5. Afterward, fetch the chain transition histories made by each FLU and verify that no transition was ever unsafe.

Behavior in symmetric network partitions

The simulator has yet to find an error. This is both really cool and really terrifying: is this really working? No, seriously, where are the bugs? Good question. Both the algorithm and the simulator need review and futher study.

In fact, it'd be awesome if I could work with someone who has more TLA+ experience than I do to work on a formal specification of the self-management algorithm and verify its correctness.

Behavior in asymmetric network partitions

The simulator's behavior during stable periods where at least one node is the victim of an asymmetric network partition is … weird, wonderful, and something I don't completely understand yet. This is another place where we need more eyes reviewing and trying to poke holes in the algorithm.

In cases where any node is a victim of an asymmetric network partition, the algorithm oscillates in a very predictable way: each node X makes the same P_newprop projection at epoch E that X made during a previous recent epoch E-delta (where delta is small, usually much less than 10). However, at least one node makes a proposal that makes rough consensus impossible. When any epoch E is not acceptable (because some node disagrees about something, e.g., which nodes are down), the result is more new rounds of proposals.

Because any node X's proposal isn't any different than X's last proposal, the system spirals into an infinite loop of never-fully-agreed-upon proposals. This is … really cool, I think.

From the sole perspective of any single participant node, the pattern of this infinite loop is easy to detect.

Were my last 2*L proposals were exactly the same? (where L is the maximum possible chain length (i.e. if all chain members are fully operational))

When detected, the local node moves to a slightly different mode of operation: it starts suspecting that a "proposal flapping" series of events is happening. (The name "flap" is taken from IP network routing, where a "flapping route" is an oscillating state of churn within the routing fabric where one or more routes change, usually in a rapid & very disruptive manner.)

If flapping is suspected, then the count of number of flap cycles is counted. If the local node sees all participants (including itself) flapping with the same relative proposed projection for 2L times in a row (where L is the maximum length of the chain), then the local node has firm evidence that there is an asymmetric network partition somewhere in the system. The pattern of proposals is analyzed, and the local node makes a decision:

  1. The local node is directly affected by the network partition. The result: stop making new projection proposals until the failure detector belives that a new status change has taken place.
  2. The local node is not directly affected by the network partition. The result: continue participating in the system by continuing new self-management algorithm iterations.

After the asymmetric partition victims have "taken themselves out of the game" temporarily, then the remaining participants rapidly converge to rough consensus and then a visibly unanimous proposal. For as long as the network remains partitioned but stable, any new iteration of the self-management algorithm stops without externally-visible effects. (I.e., it stops at the bottom of the flowchart's Column A.)

Prototype notes

Mid-March 2015

I've come to realize that the property that causes the nice property of "Were my last 2L proposals identical?" also requires that the proposals be stable. If a participant notices, "Hey, there's flapping happening, so I'll propose a different projection P_different", then the very act of proposing P_different disrupts the "last 2L proposals identical" cycle the enables us to detect flapping. We kill the goose that's laying our golden egg.

I've been working on the idea of "nested" projections, namely an "outer" and "inner" projection. Only the "outer projection" is used for cycle detection. The "inner projection" is the same as the outer projection when flapping is not detected. When flapping is detected, then the inner projection is one that excludes all nodes that the outer projection has identified as victims of asymmetric partition.

This inner projection technique may or may not work well enough to use? It would require constant flapping of the outer proposal, which is going to consume CPU and also chew up projection store keys with the flapping churn. That churn would continue as long as an asymmetric partition exists. The simplest way to cope with this would be to reduce proposal rates significantly, say 10x or 50x slower, to slow churn down to proposals from several-per-second to perhaps several-per-minute?