The Hyperledger Fabric architecture delivers the following advantages:
Part I: Elements of the architecture relevant to Hyperledger Fabricv1
System architecture
Basic workflow of transaction endorsement
Endorsement policies
Part II: Post-v1 elements of the architecture
Ledger checkpointing (pruning)
The blockchain is a distributed system consisting of many nodes thatcommunicate with each other. The blockchain runs programs calledchaincode, holds state and ledger data, and executes transactions. Thechaincode is the central element as transactions are operations invokedon the chaincode. Transactions have to be “endorsed” and only endorsedtransactions may be committed and have an effect on the state. There mayexist one or more special chaincodes for management functions andparameters, collectively called system chaincodes.
Transactions may be of two types:
As described later, deploy transactions are special cases of invoketransactions, where a deploy transaction that creates new chaincode,corresponds to an invoke transaction on a system chaincode.
Remark: This document currently assumes that a transaction eithercreates new chaincode or invokes an operation provided by *one alreadydeployed chaincode. This document does not yet describe: a)optimizations for query (read-only) transactions (included in v1), b)support for cross-chaincode transactions (post-v1 feature).*
The latest state of the blockchain (or, simply, state) is modeled as aversioned key-value store (KVS), where keys are names and values arearbitrary blobs. These entries are manipulated by the chaincodes(applications) running on the blockchain through put and getKVS-operations. The state is stored persistently and updates to thestate are logged. Notice that versioned KVS is adopted as state model,an implementation may use actual KVSs, but also RDBMSs or any othersolution.
More formally, state s is modeled as an element of a mappingK -> (V X N), where:
K is a set of keysV is a set of valuesN is an infinite ordered set of version numbers. Injectivefunction next: N -> N takes an element of N and returns thenext version number.Both V and N contain a special element ⊥ (empty type), which isin case of N the lowest element. Initially all keys are mapped to(⊥, ⊥). For s(k)=(v,ver) we denote v by s(k).value,and ver by s(k).version.
KVS operations are modeled as follows:
put(k,v) for k ∈ K and v ∈ V, takes the blockchainstate s and changes it to s' such thats'(k)=(v,next(s(k).version)) with s'(k')=s(k') for allk'!=k.get(k) returns s(k).State is maintained by peers, but not by orderers and clients.
State partitioning. Keys in the KVS can be recognized from theirname to belong to a particular chaincode, in the sense that onlytransaction of a certain chaincode may modify the keys belonging to thischaincode. In principle, any chaincode can read the keys belonging toother chaincodes. Support for cross-chaincode transactions, that modifythe state belonging to two or more chaincodes is a post-v1 feature.
Ledger provides a verifiable history of all successful state changes (wetalk about valid transactions) and unsuccessful attempts to changestate (we talk about invalid transactions), occurring during theoperation of the system.
Ledger is constructed by the ordering service (see Sec 1.3.3) as atotally ordered hashchain of blocks of (valid or invalid)transactions. The hashchain imposes the total order of blocks in aledger and each block contains an array of totally ordered transactions.This imposes total order across all transactions.
Ledger is kept at all peers and, optionally, at a subset of orderers. Inthe context of an orderer we refer to the Ledger as toOrdererLedger, whereas in the context of a peer we refer to theledger as to PeerLedger. PeerLedger differs from theOrdererLedger in that peers locally maintain a bitmask that tellsapart valid transactions from invalid ones (see Section XX for moredetails).
Peers may prune PeerLedger as described in Section XX (post-v1feature). Orderers maintain OrdererLedger for fault-tolerance andavailability (of the PeerLedger) and may decide to prune it atanytime, provided that properties of the ordering service (see Sec.1.3.3) are maintained.
The ledger allows peers to replay the history of all transactions and toreconstruct the state. Therefore, state as described in Sec 1.2.1 is anoptional datastructure.
Nodes are the communication entities of the blockchain. A “node” is onlya logical function in the sense that multiple nodes of different typescan run on the same physical server. What counts is how nodes aregrouped in “trust domains” and associated to logical entities thatcontrol them.
There are three types of nodes:
The types of nodes are explained next in more detail.
The client represents the entity that acts on behalf of an end-user. Itmust connect to a peer for communicating with the blockchain. The clientmay connect to any peer of its choice. Clients create and thereby invoketransactions.
As detailed in Section 2, clients communicate with both peers and theordering service.
A peer receives ordered state updates in the form of blocks from theordering service and maintain the state and the ledger.
Peers can additionally take up a special role of an endorsing peer,or an endorser. The special function of an endorsing peer occurswith respect to a particular chaincode and consists in endorsing atransaction before it is committed. Every chaincode may specify anendorsement policy that may refer to a set of endorsing peers. Thepolicy defines the necessary and sufficient conditions for a validtransaction endorsement (typically a set of endorsers’ signatures), asdescribed later in Sections 2 and 3. In the special case of deploytransactions that install new chaincode the (deployment) endorsementpolicy is specified as an endorsement policy of the system chaincode.
The orderers form the ordering service, i.e., a communication fabricthat provides delivery guarantees. The ordering service can beimplemented in different ways: ranging from a centralized service (usede.g., in development and testing) to distributed protocols that targetdifferent network and node fault models.
Ordering service provides a shared communication channel to clientsand peers, offering a broadcast service for messages containingtransactions. Clients connect to the channel and may broadcast messageson the channel which are then delivered to all peers. The channelsupports atomic delivery of all messages, that is, messagecommunication with total-order delivery and (implementation specific)reliability. In other words, the channel outputs the same messages toall connected peers and outputs them to all peers in the same logicalorder. This atomic communication guarantee is also called total-orderbroadcast, atomic broadcast, or consensus in the context ofdistributed systems. The communicated messages are the candidatetransactions for inclusion in the blockchain state.
Partitioning (ordering service channels). Ordering service maysupport multiple channels similar to the topics of apublish/subscribe (pub/sub) messaging system. Clients can connect to agiven channel and can then send messages and obtain the messages thatarrive. Channels can be thought of as partitions - clients connecting toone channel are unaware of the existence of other channels, but clientsmay connect to multiple channels. Even though some ordering serviceimplementations included with Hyperledger Fabric support multiplechannels, for simplicity of presentation, in the rest of thisdocument, we assume ordering service consists of a single channel/topic.
Ordering service API. Peers connect to the channel provided by theordering service, via the interface provided by the ordering service.The ordering service API consists of two basic operations (moregenerally asynchronous events):
TODO add the part of the API for fetching particular blocks underclient/peer specified sequence numbers.
broadcast(blob): a client calls this to broadcast an arbitrarymessage blob for dissemination over the channel. This is alsocalled request(blob) in the BFT context, when sending a requestto a service.deliver(seqno, prevhash, blob): the ordering service calls thison the peer to deliver the message blob with the specifiednon-negative integer sequence number (seqno) and hash of the mostrecently delivered blob (prevhash). In other words, it is anoutput event from the ordering service. deliver() is alsosometimes called notify() in pub-sub systems or commit() inBFT systems.Ledger and block formation. The ledger (see also Sec. 1.2.2)contains all data output by the ordering service. In a nutshell, it is asequence of deliver(seqno, prevhash, blob) events, which form a hashchain according to the computation of prevhash described before.
Most of the time, for efficiency reasons, instead of outputtingindividual transactions (blobs), the ordering service will group (batch)the blobs and output blocks within a single deliver event. In thiscase, the ordering service must impose and convey a deterministicordering of the blobs within each block. The number of blobs in a blockmay be chosen dynamically by an ordering service implementation.
In the following, for ease of presentation, we define ordering serviceproperties (rest of this subsection) and explain the workflow oftransaction endorsement (Section 2) assuming one blob per deliverevent. These are easily extended to blocks, assuming that a deliverevent for a block corresponds to a sequence of individual deliverevents for each blob within a block, according to the above mentioneddeterministic ordering of blobs within a block.
Ordering service properties
The guarantees of the ordering service (or atomic-broadcast channel)stipulate what happens to a broadcasted message and what relations existamong delivered messages. These guarantees are as follows:
Safety (consistency guarantees): As long as peers are connectedfor sufficiently long periods of time to the channel (they candisconnect or crash, but will restart and reconnect), they will seean identical series of delivered (seqno, prevhash, blob)messages. This means the outputs (deliver() events) occur in thesame order on all peers and according to sequence number and carryidentical content (blob and prevhash) for the same sequencenumber. Note this is only a logical order, and adeliver(seqno, prevhash, blob) on one peer is not required tooccur in any real-time relation to deliver(seqno, prevhash, blob)that outputs the same message at another peer. Put differently, givena particular seqno, no two correct peers deliver differentprevhash or blob values. Moreover, no value blob isdelivered unless some client (peer) actually calledbroadcast(blob) and, preferably, every broadcasted blob is onlydelivered once.
Furthermore, the deliver() event contains the cryptographic hashof the data in the previous deliver() event (prevhash). Whenthe ordering service implements atomic broadcast guarantees,prevhash is the cryptographic hash of the parameters from thedeliver() event with sequence number seqno-1. Thisestablishes a hash chain across deliver() events, which is usedto help verify the integrity of the ordering service output, asdiscussed in Sections 4 and 5 later. In the special case of the firstdeliver() event, prevhash has a default value.
Liveness (delivery guarantee): Liveness guarantees of theordering service are specified by a ordering service implementation.The exact guarantees may depend on the network and node fault model.
In principle, if the submitting client does not fail, the orderingservice should guarantee that every correct peer that connects to theordering service eventually delivers every submitted transaction.
To summarize, the ordering service ensures the following properties:
deliver(seqno, prevhash0, blob0) anddeliver(seqno, prevhash1, blob1) with the same seqno,prevhash0==prevhash1 and blob0==blob1;deliver(seqno-1, prevhash0, blob0) anddeliver(seqno, prevhash, blob),prevhash = HASH(seqno-1||prevhash0||blob0).deliver(seqno, prevhash, blob) at a correct peer p, such thatseqno>0, then p already delivered an eventdeliver(seqno-1, prevhash0, blob0).deliver(seqno, prevhash, blob) at acorrect peer must be preceded by a broadcast(blob) event at some(possibly distinct) peer;broadcast(blob) and broadcast(blob'), when two eventsdeliver(seqno0, prevhash0, blob) anddeliver(seqno1, prevhash1, blob') occur at correct peers andblob == blob', then seqno0==seqno1 andprevhash0==prevhash1.broadcast(blob)then every correct peer “eventually” issues an eventdeliver(*, *, blob), where * denotes an arbitrary value.In the following we outline the high-level request flow for atransaction.
Remark: Notice that the following protocol *does not assume thatall transactions are deterministic, i.e., it allows fornon-deterministic transactions.*
To invoke a transaction, the client sends a PROPOSE message to a setof endorsing peers of its choice (possibly not at the same time - seeSections 2.1.2. and 2.3.). The set of endorsing peers for a givenchaincodeID is made available to client via peer, which in turnknows the set of endorsing peers from endorsement policy (see Section3). For example, the transaction could be sent to all endorsers of agiven chaincodeID. That said, some endorsers could be offline,others may object and choose not to endorse the transaction. Thesubmitting client tries to satisfy the policy expression with theendorsers available.
In the following, we first detail PROPOSE message format and thendiscuss possible patterns of interaction between submitting client andendorsers.
PROPOSE message formatThe format of a PROPOSE message is tx is a mandatory and anchor optional argument explained in thefollowing.
tx=, where
clientID is an ID of the submitting client,chaincodeID refers to the chaincode to which the transactionpertains,txPayload is the payload containing the submitted transactionitself,timestamp is a monotonically increasing (for every newtransaction) integer maintained by the client,clientSig is signature of a client on other fields of tx.The details of txPayload will differ between invoke transactionsand deploy transactions (i.e., invoke transactions referring to adeploy-specific system chaincode). For an invoke transaction,txPayload would consist of two fields
txPayload = metadata> , where
operation denotes the chaincode operation (function) andarguments,metadata denotes attributes related to the invocation.For a deploy transaction, txPayload would consist of threefields
txPayload = metadata, policies> , where
source denotes the source code of the chaincode,metadata denotes attributes related to the chaincode andapplication,policies contains policies related to the chaincode thatare accessible to all peers, such as the endorsement policy.Note that endorsement policies are not supplied withtxPayload in a deploy transaction, buttxPayload of a deploy contains endorsement policy ID andits parameters (see Section 3).anchor contains read version dependencies, or morespecifically, key-version pairs (i.e., anchor is a subset ofKxN), that binds or “anchors” the PROPOSE request tospecified versions of keys in a KVS (see Section 1.2.). If the clientspecifies the anchor argument, an endorser endorses a transactiononly upon read version numbers of corresponding keys in its localKVS match anchor (see Section 2.2. for more details).
Cryptographic hash of tx is used by all nodes as a uniquetransaction identifier tid (i.e., tid=HASH(tx)). The clientstores tid in memory and waits for responses from endorsing peers.
The client decides on the sequence of interaction with endorsers. Forexample, a client would typically send anchor argument) to a single endorser, which would then producethe version dependencies (anchor) which the client can later on useas an argument of its PROPOSE message to other endorsers. As anotherexample, the client could directly send anchor) to all endorsers of its choice. Different patterns ofcommunication are possible and client is free to decide on those (seealso Section 2.3.).
On reception of a epID first verifies the client’s signatureclientSig and then simulates a transaction. If the client specifiesanchor then endorsing peer simulates the transactions only upon readversion numbers (i.e., readset as defined below) of correspondingkeys in its local KVS match those version numbers specified byanchor.
Simulating a transaction involves endorsing peer tentatively executinga transaction (txPayload), by invoking the chaincode to which thetransaction refers (chaincodeID) and the copy of the state that theendorsing peer locally holds.
As a result of the execution, the endorsing peer computes read versiondependencies (readset) and state updates (writeset), alsocalled MVCC+postimage info in DB language.
Recall that the state consists of key-value pairs. All key-value entriesare versioned; that is, every entry contains ordered versioninformation, which is incremented each time the value stored undera key is updated. The peer that interprets the transaction records allkey-value pairs accessed by the chaincode, either for reading or for writing,but the peer does not yet update its state. More specifically:
s before an endorsing peer executes a transaction,for every key k read by the transaction, pair(k,s(k).version) is added to readset.k modified by the transaction to thenew value v', pair (k,v') is added to writeset.Alternatively, v' could be the delta of the new value to previousvalue (s(k).value).If a client specifies anchor in the PROPOSE message then clientspecified anchor must equal readset produced by endorsing peerwhen simulating the transaction.
Then, the peer forwards internally tran-proposal (and possiblytx) to the part of its (peer’s) logic that endorses a transaction,referred to as endorsing logic. By default, endorsing logic at apeer accepts the tran-proposal and simply signs thetran-proposal. However, endorsing logic may interpret arbitraryfunctionality, to, e.g., interact with legacy systems withtran-proposal and tx as inputs to reach the decision whether toendorse a transaction or not.
If endorsing logic decides to endorse a transaction, it sendstx.clientID), where:
tran-proposal := (epID,tid,chaincodeID,txContentBlob,readset,writeset),
where txContentBlob is chaincode/transaction specificinformation. The intention is to have txContentBlob used as somerepresentation of tx (e.g., txCOntentBlob=tx.txPayload).
epSig is the endorsing peer’s signature on tran-proposal
Else, in case the endorsing logic refuses to endorse the transaction, anendorser may send a message (TRANSACTION-INVALID, tid, REJECTED)to the submitting client.
Notice that an endorser does not change its state in this step, theupdates produced by transaction simulation in the context of endorsementdo not affect the state!
The submitting client waits until it receives “enough” messages andsignatures on (TRANSACTION-ENDORSED, tid, *, *) statements toconclude that the transaction proposal is endorsed. As discussed inSection 2.1.2., this may involve one or more round-trips of interactionwith endorsers.
The exact number of “enough” depend on the chaincode endorsement policy(see also Section 3). If the endorsement policy is satisfied, thetransaction has been endorsed; note that it is not yet committed. Thecollection of signed TRANSACTION-ENDORSED messages from endorsingpeers which establish that a transaction is endorsed is called anendorsement and denoted by endorsement.
If the submitting client does not manage to collect an endorsement for atransaction proposal, it abandons this transaction with an option toretry later.
For transaction with a valid endorsement, we now start using theordering service. The submitting client invokes ordering service usingthe broadcast(blob), where blob=endorsement. If the client doesnot have capability of invoking ordering service directly, it may proxyits broadcast through some peer of its choice. Such a peer must betrusted by the client not to remove any message from the endorsementor otherwise the transaction may be deemed invalid. Notice that,however, a proxy peer may not fabricate a valid endorsement.
When an event deliver(seqno, prevhash, blob) occurs and a peer hasapplied all state updates for blobs with sequence number lower thanseqno, a peer does the following:
blob.endorsement is valid according to thepolicy of the chaincode (blob.tran-proposal.chaincodeID) to whichit refers.blob.endorsement.tran-proposal.readset) have not been violatedmeanwhile. In more complex use cases, tran-proposal fields inendorsement may differ and in this case endorsement policy (Section3) specifies how the state evolves.Verification of dependencies can be implemented in different ways,according to a consistency property or “isolation guarantee” that ischosen for the state updates. Serializability is a default isolationguarantee, unless chaincode endorsement policy specifies a differentone. Serializability can be provided by requiring the version associatedwith every key in the readset to be equal to that key’s version inthe state, and rejecting transactions that do not satisfy thisrequirement.
PeerLedger, appliesblob.endorsement.tran-proposal.writeset to blockchain state (iftran-proposals are the same, otherwise endorsement policy logicdefines the function that takes blob.endorsement).blob.endorsement fails,the transaction is invalid and the peer marks the transaction with 0in the bitmask of the PeerLedger. It is important to note thatinvalid transactions do not change the state.Note that this is sufficient to have all (correct) peers have the samestate after processing a deliver event (block) with a given sequencenumber. Namely, by the guarantees of the ordering service, all correctpeers will receive an identical sequence ofdeliver(seqno, prevhash, blob) events. As the evaluation of theendorsement policy and evaluation of version dependencies in readsetare deterministic, all correct peers will also come to the sameconclusion whether a transaction contained in a blob is valid. Hence,all peers commit and apply the same sequence of transactions and updatetheir state in the same way.
Figure 1. Illustration of one possible transaction flow (common-case path).
An endorsement policy, is a condition on what endorses atransaction. Blockchain peers have a pre-specified set of endorsementpolicies, which are referenced by a deploy transaction that installsspecific chaincode. Endorsement policies can be parametrized, and theseparameters can be specified by a deploy transaction.
To guarantee blockchain and security properties, the set of endorsementpolicies should be a set of proven policies with limited set offunctions in order to ensure bounded execution time (termination),determinism, performance and security guarantees.
Dynamic addition of endorsement policies (e.g., by deploytransaction on chaincode deploy time) is very sensitive in terms ofbounded policy evaluation time (termination), determinism, performanceand security guarantees. Therefore, dynamic addition of endorsementpolicies is not allowed, but can be supported in future.
A transaction is declared valid only if it has been endorsed accordingto the policy. An invoke transaction for a chaincode will first have toobtain an endorsement that satisfies the chaincode’s policy or it willnot be committed. This takes place through the interaction between thesubmitting client and endorsing peers as explained in Section 2.
Formally the endorsement policy is a predicate on the endorsement, andpotentially further state that evaluates to TRUE or FALSE. For deploytransactions the endorsement is obtained according to a system-widepolicy (for example, from the system chaincode).
An endorsement policy predicate refers to certain variables. Potentiallyit may refer to:
endorsement and endorsement.tran-proposal;The above list is ordered by increasing expressiveness and complexity,that is, it will be relatively simple to support policies that onlyrefer to keys and identities of nodes.
The evaluation of an endorsement policy predicate must bedeterministic. An endorsement shall be evaluated locally by every peersuch that a peer does not need to interact with other peers, yet allcorrect peers evaluate the endorsement policy in the same way.
The predicate may contain logical expressions and evaluates to TRUE orFALSE. Typically the condition will use digital signatures on thetransaction invocation issued by endorsing peers for the chaincode.
Suppose the chaincode specifies the endorser setE = {Alice, Bob, Charlie, Dave, Eve, Frank, George}. Some examplepolicies:
tran-proposal from all membersof E.tran-proposal from endorsing peersaccording to the condition(Alice OR Bob) AND (any two of: Charlie, Dave, Eve, Frank, George).tran-proposal by any 5 out of the 7endorsers. (More generally, for chaincode with n > 3f endorsers,valid signatures by any 2f+1 out of the n endorsers, or byany group of more than (n+f)/2 endorsers.){Alice=49, Bob=15, Charlie=15, Dave=10, Eve=7, Frank=3, George=1},where the total stake is 100: The policy requires valid signaturesfrom a set that has a majority of the stake (i.e., a group withcombined stake strictly more than 50), such as {Alice, X} withany X different from George, or{everyone together except Alice}. And so on.tran-proposal1 and validsignatures from (any two of: Charlie, Dave, Eve, Frank, George)on tran-proposal2, where tran-proposal1 andtran-proposal2 differ only in their endorsing peers and stateupdates.How useful these policies are will depend on the application, on thedesired resilience of the solution against failures or misbehavior ofendorsers, and on various other properties.
PeerLedger checkpointing (pruning)To maintain the abstraction of a ledger that contains only valid andcommitted transactions (that appears in Bitcoin, for example), peersmay, in addition to state and Ledger, maintain the Validated Ledger (orVLedger). This is a hash chain derived from the ledger by filtering outinvalid transactions.
The construction of the VLedger blocks (called here vBlocks) proceedsas follows. As the PeerLedger blocks may contain invalidtransactions (i.e., transactions with invalid endorsement or withinvalid version dependencies), such transactions are filtered out bypeers before a transaction from a block becomes added to a vBlock. Everypeer does this by itself (e.g., by using the bitmask associated withPeerLedger). A vBlock is defined as a block without the invalidtransactions, that have been filtered out. Such vBlocks are inherentlydynamic in size and may be empty. An illustration of vBlock constructionis given in the figure below.
Figure 2. Illustration of validated ledger block (vBlock) formation from ledger (PeerLedger) blocks.
vBlocks are chained together to a hash chain by every peer. Morespecifically, every block of a validated ledger contains:
PeerLedger) from whichthe current vBlock is derived.All this information is concatenated and hashed by a peer, producing thehash of the vBlock in the validated ledger.
PeerLedger CheckpointingThe ledger contains invalid transactions, which may not necessarily berecorded forever. However, peers cannot simply discard PeerLedgerblocks and thereby prune PeerLedger once they establish thecorresponding vBlocks. Namely, in this case, if a new peer joins thenetwork, other peers could not transfer the discarded blocks (pertainingto PeerLedger) to the joining peer, nor convince the joining peer ofthe validity of their vBlocks.
To facilitate pruning of the PeerLedger, this document describes acheckpointing mechanism. This mechanism establishes the validity ofthe vBlocks across the peer network and allows checkpointed vBlocks toreplace the discarded PeerLedger blocks. This, in turn, reducesstorage space, as there is no need to store invalid transactions. Italso reduces the work to reconstruct the state for new peers that jointhe network (as they do not need to establish validity of individualtransactions when reconstructing the state by replaying PeerLedger,but may simply replay the state updates contained in the validatedledger).
Checkpointing is performed periodically by the peers every CHK blocks,where CHK is a configurable parameter. To initiate a checkpoint, thepeers broadcast (e.g., gossip) to other peers messageblockno is the current blocknumber and blocknohash is itsrespective hash, stateHash is the hash of the latest state (producedby e.g., a Merkle hash) upon validation of block blockno andpeerSig is peer’s signature on(CHECKPOINT,blocknohash,blockno,stateHash), referring to thevalidated ledger.
A peer collects CHECKPOINT messages until it obtains enoughcorrectly signed messages with matching blockno, blocknohash andstateHash to establish a valid checkpoint (see Section 4.2.2.).
Upon establishing a valid checkpoint for block number blockno withblocknohash, a peer:
blockno>latestValidCheckpoint.blockno, then a peer assignslatestValidCheckpoint=(blocknohash,blockno),latestValidCheckpointProof,stateHash tolatestValidCheckpointedState,PeerLedger up to block number blockno(inclusive).Clearly, the checkpointing protocol raises the following questions:When can a peer prune its ``PeerLedger``? How many ``CHECKPOINT``messages are “sufficiently many”?. This is defined by a checkpointvalidity policy, with (at least) two possible approaches, which mayalso be combined:
CHECKPOINT messages are sufficient to establisha valid checkpoint. For example, LCVP at peer Alice may define thatAlice needs to receive CHECKPOINT message from Bob, or fromboth Charlie and Dave.
京公网安备 11010802041100号 | 京ICP备19059560号-4 | PHP1.CN 第一PHP社区 版权所有