github.com/susy-go/susy-graviton@v0.0.0-20190614130430-36cddae42305/swarm/network/README.md (about) 1 ## Streaming 2 3 Streaming is a new protocol of the swarm bzz bundle of protocols. 4 This protocol provides the basic logic for chunk-based data flow. 5 It implements simple retrieve requests and delivery using priority queue. 6 A data exchange stream is a directional flow of chunks between peers. 7 The source of datachunks is the upstream, the receiver is called the 8 downstream peer. Each streaming protocol defines an outgoing streamer 9 and an incoming streamer, the former installing on the upstream, 10 the latter on the downstream peer. 11 12 Subscribe on StreamerPeer launches an incoming streamer that sends 13 a subscribe msg upstream. The streamer on the upstream peer 14 handles the subscribe msg by installing the relevant outgoing streamer 15 . The modules now engage in a process of upstream sending a sequence of hashes of 16 chunks downstream (OfferedHashesMsg). The downstream peer evaluates which hashes are needed 17 and get it delivered by sending back a msg (WantedHashesMsg). 18 19 Historical syncing is supported - currently not the right abstraction -- 20 state kept across sessions by saving a series of intervals after their last 21 batch actually arrived. 22 23 Live streaming is also supported, by starting session from the first item 24 after the subscription. 25 26 Provable data exchange. In case a stream represents a swarm document's data layer 27 or higher level chunks, streaming up to a certain index is always provable. It saves on 28 sending intermediate chunks. 29 30 Using the streamer logic, various stream types are easy to implement: 31 32 * light node requests: 33 * url lookup with offset 34 * document download 35 * document upload 36 * syncing 37 * live session syncing 38 * historical syncing 39 * simple retrieve requests and deliveries 40 * swarm feeds streams 41 * receipting for finger pointing 42 43 ## Syncing 44 45 Syncing is the process that makes sure storer nodes end up storing all and only the chunks that are requested from them. 46 47 ### Requirements 48 49 - eventual consistency: so each chunk historical should be syncable 50 - since the same chunk can and will arrive from many peers, (network traffic should be 51 optimised, only one transfer of data per chunk) 52 - explicit request deliveries should be prioritised higher than recent chunks received 53 during the ongoing session which in turn should be higher than historical chunks. 54 - insured chunks should get receipted for finger pointing litigation, the receipts storage 55 should be organised efficiently, upstream peer should also be able to find these 56 receipts for a deleted chunk easily to refute their challenge. 57 - syncing should be resilient to cut connections, metadata should be persisted that 58 keep track of syncing state across sessions, historical syncing state should survive restart 59 - extra data structures to support syncing should be kept at minimum 60 - syncing is not organized separately for chunk types (Swarm feed updates v regular content chunk) 61 - various types of streams should have common logic abstracted 62 63 Syncing is now entirely mediated by the localstore, ie., no processes or memory leaks due to network contention. 64 When a new chunk is stored, its chunk hash is index by proximity bin 65 66 peers syncronise by getting the chunks closer to the downstream peer than to the upstream one. 67 Consequently peers just sync all stored items for the kad bin the receiving peer falls into. 68 The special case of nearest neighbour sets is handled by the downstream peer 69 indicating they want to sync all kademlia bins with proximity equal to or higher 70 than their depth. 71 72 This sync state represents the initial state of a sync connection session. 73 Retrieval is dictated by downstream peers simply using a special streamer protocol. 74 75 Syncing chunks created during the session by the upstream peer is called live session syncing 76 while syncing of earlier chunks is historical syncing. 77 78 Once the relevant chunk is retrieved, downstream peer looks up all hash segments in its localstore 79 and sends to the upstream peer a message with a a bitvector to indicate 80 missing chunks (e.g., for chunk `k`, hash with chunk internal index which case ) 81 new items. In turn upstream peer sends the relevant chunk data alongside their index. 82 83 On sending chunks there is a priority queue system. If during looking up hashes in its localstore, 84 downstream peer hits on an open request then a retrieve request is sent immediately to the upstream peer indicating 85 that no extra round of checks is needed. If another peers syncer hits the same open request, it is slightly unsafe to not ask 86 that peer too: if the first one disconnects before delivering or fails to deliver and therefore gets 87 disconnected, we should still be able to continue with the other. The minimum redundant traffic coming from such simultaneous 88 eventualities should be sufficiently rare not to warrant more complex treatment. 89 90 Session syncing involves downstream peer to request a new state on a bin from upstream. 91 using the new state, the range (of chunks) between the previous state and the new one are retrieved 92 and chunks are requested identical to the historical case. After receiving all the missing chunks 93 from the new hashes, downstream peer will request a new range. If this happens before upstream peer updates a new state, 94 we say that session syncing is live or the two peers are in sync. In general the time interval passed since downstream peer request up to the current session cursor is a good indication of a permanent (probably increasing) lag. 95 96 If there is no historical backlog, and downstream peer has an acceptable 'last synced' tag, then it is said to be fully synced with the upstream peer. 97 If a peer is fully synced with all its storer peers, it can advertise itself as globally fully synced. 98 99 The downstream peer persists the record of the last synced offset. When the two peers disconnect and 100 reconnect syncing can start from there. 101 This situation however can also happen while historical syncing is not yet complete. 102 Effectively this means that the peer needs to persist a record of an arbitrary array of offset ranges covered. 103 104 ### Delivery requests 105 106 once the appropriate ranges of the hashstream are retrieved and buffered, downstream peer just scans the hashes, looks them up in localstore, if not found, create a request entry. 107 The range is referenced by the chunk index. Alongside the name (indicating the stream, e.g., content chunks for bin 6) and the range 108 downstream peer sends a 128 long bitvector indicating which chunks are needed. 109 Newly created requests are satisfied bound togsophy in a waitgroup which when done, will promptt sending the next one. 110 to be able to do check and storage concurrently, we keep a buffer of one, we start with two batches of hashes. 111 If there is nothing to give, upstream peers SetNextBatch is blocking. Subscription ends with an unsubscribe. which removes the syncer from the map. 112 113 Canceling requests (for instance the late chunks of an erasure batch) should be a chan closed 114 on the request 115 116 Simple request is also a subscribe 117 different streaming protocols are different p2p protocols with same message types. 118 the constructor is the Run function itself. which takes a streamerpeer as argument 119 120 121 ### provable streams 122 123 The swarm hash over the hash stream has many advantages. It implements a provable data transfer 124 and provide efficient storage for receipts in the form of inclusion proofs useable for finger pointing litigation. 125 When challenged on a missing chunk, upstream peer will provide an inclusion proof of a chunk hash against the state of the 126 sync stream. In order to be able to generate such an inclusion proof, upstream peer needs to store the hash index (counting consecutive hash-size segments) alongside the chunk data and preserve it even when the chunk data is deleted until the chunk is no longer insured. 127 if there is no valid insurance on the files the entry may be deleted. 128 As long as the chunk is preserved, no takeover proof will be needed since the node can respond to any challenge. 129 However, once the node needs to delete an insured chunk for capacity reasons, a receipt should be available to 130 refute the challenge by finger pointing to a downstream peer. 131 As part of the deletion protocol then, hashes of insured chunks to be removed are pushed to an infinite stream for every bin. 132 133 Downstream peer on the other hand needs to make sure that they can only be finger pointed about a chunk they did receive and store. 134 For this the check of a state should be exhaustive. If historical syncing finishes on one state, all hashes before are covered, no 135 surprises. In other words historical syncing this process is self verifying. With session syncing however, it is not enough to check going back covering the range from old offset to new. Continuity (i.e., that the new state is extension of the old) needs to be verified: after downstream peer reads the range into a buffer, it appends the buffer the last known state at the last known offset and verifies the resulting hash matches 136 the latest state. Past intervals of historical syncing are checked via the session root. 137 Upstream peer signs the states, downstream peers can use as handover proofs. 138 Downstream peers sign off on a state togsophy with an initial offset. 139 140 Once historical syncing is complete and the session does not lag, downstream peer only preserves the latest upstream state and store the signed version. 141 142 Upstream peer needs to keep the latest takeover states: each deleted chunk's hash should be covered by takeover proof of at least one peer. If historical syncing is complete, upstream peer typically will store only the latest takeover proof from downstream peer. 143 Crucially, the structure is totally independent of the number of peers in the bin, so it scales extremely well. 144 145 ## implementation 146 147 The simplest protocol just involves upstream peer to prefix the key with the kademlia proximity order (say 0-15 or 0-31) 148 and simply iterate on index per bin when syncing with a peer. 149 150 priority queues are used for sending chunks so that user triggered requests should be responded to first, session syncing second, and historical with lower priority. 151 The request on chunks remains implemented as a dataless entry in the memory store. 152 The lifecycle of this object should be more carefully thought through, ie., when it fails to retrieve it should be removed.