5. Decentralized Oracles With Fluence And Aqua

Overview
An oracle is some device that provides real-world, off-chain data to deterministic on-chain consumers such as a smart contract. A decentralized oracle draws from multiple, purportedly (roughly) equal input sources to minimize or even eliminate single source pitfalls such as man-in-the-middle attacks(MITM) or provider manipulation. For example, a decentralized price oracle for, say, ETH/USD, could poll several DEXs for ETH/USD prices. Since smart contracts, especially those deployed on EVMs, can't directly call off-chain resources, oracles play a critical "middleware" role in the decentralized, trustless ecosystem. See Figure 1.
Unlike single source oracles, multi-source oracles require some consensus mechanism to convert multiple input sources over the same target parameter into reliable point or range data suitable for third party, e.g., smart contract, consumption. Such "consensus over inputs" may take the form of simple summary statistics, e.g., mean, or one of many other methods.
Given the importance of oracles to the Web3 ecosystem, it's not surprising to see a variety of third party solutions supporting various blockchain protocols. Fluence does not provide an oracle solution per se but provides a peer-to-peer platform, tools and components for developers to quickly and easily program and compose reusable distributed data acquisition, processing and delivery services into decentralized oracle applications.
For the remainder of this section, we work through the process of developing a decentralized, multi-source timestamp oracle comprised of data acquisition, processing and delivery.
Creating A Decentralized Timestamp Oracle
Time, often in form of timestamps, plays a critical role in a large number of Web2 and Web3 applications including off-chain voting applications and on-chain clocks. Our goal is to provide a consensus timestamp sourced from multiple input sources and implement an acceptable input aggregation and processing service to arrive at either a timestamp point or range value(s).
In case you haven't set up your development environment, follow the setup instructions and clone the examples repo:
1
git clone https://github.com/fluencelabs/examples
Copied!
Timestamp Acquisition
Each Fluence peer, i.e. node in the Fluence peer-to-peer network, has the ability to provide a timestamp from a builtin service. In Aqua, we can call a timestamp function with the desired granularity, i.e., seconds or milliseconds for further processing:
1
    -- aqua timestamp sourcing
2
    on peer:
3
        ts_ms_result <- peer.timestamp_ms()
4
        -- or
5
        ts_sec_result <- peer.timestamp_sec()
6
    -- ...
Copied!
In order to decentralize our timestamp oracle, we want to poll multiple peers in the Fluence network:
1
    -- multi-peer timestamp sourcing
2
    -- ...
3
    results: *u64
4
    for peer <- many_peers_list par:
5
        on peer:
6
            results <- peer.timestamp_ms()
7
    -- ...
Copied!
In the above example, we have a list of peers and retrieve a timestamp value from each one. Note that we are polling nodes for timestamps in parallel in order to optimize toward uniformity and to collect responses in the stream variable results. See Figure 2.
The last thing to pin down concerning our timestamp acquisition is which peers to query. One possibility is to specify the peer ids of a set of desired peers to query. Alternatively, we can tap into the Kademlia neighborhood of a peer, which is a set of peers that are closest to our peer based on the XOR distance of the peer ids. Luckily, there is a builtin service we can call from Aqua that returns up to 20 neighboring peers:
1
    -- timestamps from Kademlia neighborhood
2
    results: *u64
3
    on node:
4
    k <- Op.string_to_b58(node)
5
    nodes <- Kademlia.neighborhood(k, nil, nil)
6
    for node <- nodes par:
7
      on node:
8
        try:
9
          results <- node.timestamp_ms()
10
    -- ...
Copied!
Timestamp Processing
Once we have our multiple timestamp values, we need to process them into a point or range value(s) to be useful. Whatever our processing/consensus algorithm is, we can implement it in Marine as one or more reusable, distributed services.
Fpyor example, we can rely on summary statistics and implement basic averaging to arrive at a point estimate:
1
    // ...
2
​
3
    #[marine]
4
    pub fn ts_avg(timestamps: Vec<u64>) -> f64 {
5
        timestamps.iter().sum::<u64>() as f64 / timestamps.len() as f64
6
}
7
    // ...
Copied!
Using the average to arrive at a point-estimate is simply a stake in the ground to illustrate what's possible. Actual processing algorithms may vary and, depending on a developers target audience, different algorithms may be used for different delivery targets. And Aqua makes it easy to customize workflows while emphasizing reuse.
Putting It All Together
Let's put it all together by sourcing timestamps from the Kademlia neighborhood and processing the timestamps into a consensus value. Instead of one of the summary statistics, we employ a simple, consensus algorithm that randomly selects one of the provided timestamps and then calculates a consensus score from the remaining n -1 timestamps:
1
// src.main.rs
2
// 
3
// simple consensus from timestamps
4
// params:
5
//     timestamps, u64, [0, u64_max]
6
//     tolerance, u32,  [0, u32_max]
7
//     threshold, f64, [0.0, 1.0]
8
// 1. Remove a randomly selected timestamp from the array of timestamps, ts
9
// 2. Count the number of timestamps left in the array that are withn +/- tolerance (where tolerance may be zero)
10
// 3. compare the suporting number of times stamps divided by th enumber of remaining timestamps to the threshold. if >=, consensus for selected timestamp is true else false
11
//
12
[marine]
13
fn ts_frequency(mut timestamps: Vec<u64>, tolerance: u32, threshold: f64, err_value: u64) -> Consensus {
14
    timestamps.retain(|&ts| ts != err_value);
15
    if timestamps.len() == 0 {
16
        return Consensus {
17
            err_str: "Array must have at least one element".to_string(),
18
            ..<_>::default()
19
        };
20
    }
21
​
22
    if timestamps.len() == 1 {
23
        return Consensus {
24
            n: 1,
25
            consensus_ts: timestamps[0],
26
            consensus: true,
27
            support: 1,
28
            ..<_>::default()
29
        };
30
    }
31
​
32
    if threshold < 0f64 || threshold > 1f64 {
33
        return Consensus {
34
            err_str: "Threshold needs to be between [0.0,1.0]".to_string(),
35
            ..<_>::default()
36
        };
37
    }
38
​
39
    let rnd_seed: u64 = timestamps.iter().sum();
40
    let mut rng = WyRand::new_seed(rnd_seed);
41
    let rnd_idx = rng.generate_range(0..timestamps.len());
42
    let consensus_ts = timestamps.swap_remove(rnd_idx);
43
    let mut support: u32 = 0;
44
    for ts in timestamps.iter() {
45
        if ts <= &(consensus_ts + tolerance as u64) && ts >= &(consensus_ts - tolerance as u64) {
46
            support += 1;
47
        }
48
    }
49
​
50
    let mut consensus = false;
51
    if (support as f64 / timestamps.len() as f64) >= threshold {
52
        consensus = true;
53
    }
54
​
55
    Consensus {
56
        n: timestamps.len() as u32,
57
        consensus_ts,
58
        consensus,
59
        support,
60
        err_str: "".to_string(),
61
    }
62
}
Copied!
We compile our consensus module with ./scripts/build.sh, which allows us to run the unit tests using the Wasm module with cargo +nightly test:
1
# src.main.rs
2
running 10 tests
3
test tests::ts_validation_good_consensus_false ... ok
4
test tests::test_err_val ... ok
5
test tests::test_mean_fail ... ok
6
test tests::ts_validation_good_consensus ... ok
7
test tests::ts_validation_bad_empty ... ok
8
test tests::ts_validation_good_consensus_true ... ok
9
test tests::ts_validation_good_no_support ... ok
10
test tests::test_mean_good ... ok
11
test tests::ts_validation_good_no_consensus ... ok
12
test tests::ts_validation_good_one ... ok
13
​
14
test result: ok. 10 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 18.75s
Copied!
We can now interact with our module with the Marine REPL mrepl configs/Config.toml:
1
Welcome to the Marine REPL (version 0.9.1)
2
Minimal supported versions
3
  sdk: 0.6.0
4
  interface-types: 0.20.0
5
​
6
app service was created with service id = 520a092b-85ef-43c1-9c12-444274ba2cb7
7
elapsed time 62.893047ms
8
​
9
1> i
10
Loaded modules interface:
11
data Consensus:
12
  n: u32
13
  reference_ts: u64
14
  support: u32
15
  err_str: string
16
data Oracle:
17
  n: u32
18
  avg: f64
19
  err_str: string
20
​
21
ts_oracle:
22
  fn ts_avg(timestamps: []u64, min_points: u32) -> Oracle
23
  fn ts_frequency(timestamps: []u64, tolerance: u32) -> Consensus
24
​
25
2> call ts_oracle ts_frequency [[1637182263,1637182264,1637182265,163718226,1637182266], 0, 0.66, 0]
26
result: Object({"consensus": Bool(false), "consensus_ts": Number(1637182264), "err_str": String(""), "n": Number(4), "support": Number(0)})
27
 elapsed time: 167.078µss
28
​
29
3>  call ts_oracle ts_frequency [[1637182263,1637182264,1637182265,163718226,1637182266], 5, 0.66, 0]
30
result: Object({"consensus": Bool(true), "consensus_ts": Number(1637182264), "err_str": String(""), "n": Number(4), "support": Number(3)})
31
 elapsed time: 63.291µs
Copied!
In our first call at prompt 2>, we set a tolerance of 0 and, given our array of timestamps, have no support for the chosen timestamps, whereas in the next call, 3>,we increase the tolerance parameter and obtain a consensus result.
All looks satisfactory and we are ready to deploy our module with ./scripts/deploy.sh, which write-appends the deployment response data, including the service id, to a local file named deployed_service.data:
1
client seed: 7UNmJPMWdLmrwAtGrpJXNrrcK7tEZHCjvKbGdSzEizEr
2
client peerId: 12D3KooWBeEsUnMV9MZ6QGMfaxcTvw8mGFEMGDe7rhKN8RQv1Gs8
3
relay peerId: 12D3KooWFEwNWcHqi9rtsmDhsYcDbRUCDXH84RC4FW6UfsFWaoHi
4
service id: 61a86f67-ffc2-4dea-8746-fd4f04d9c75b
5
service created successfully
Copied!
With the service in place, let's have a look at our Aqua script. Recall, we want to poll the Kademlia neighborhood for timestamps and then call the ts_oracle method of our service with the array of timestamps and tolerance parameters as well as the (peer id, service id) parameters of our deployed service:
1
-- aqua/ts_oracle.aqua
2
-- <snip>
3
​
4
func ts_oracle_with_consensus(tolerance: u32, threshold: f64, err_value:u64, node:string, oracle_service_id:string)-> Consensus, []string:
5
  rtt  = 1000                                   
6
  res: *u64                                       -- 4
7
  msg = "timeout"
8
  dead_peers: *string
9
  on node:
10
    k <- Op.string_to_b58(node)
11
    nodes <- Kademlia.neighborhood(k, nil, nil)   -- 1
12
    for n <- nodes par:                           -- 3
13
        status: *string
14
        on n:                                     -- 7
15
            res <- Peer.timestamp_ms()            -- 2
16
            status <<- "success"                  -- 9
17
        par status <- Peer.timeout(rtt, msg)      -- 8
18
        if status! != "success":
19
          res <<- err_value                       -- 10 
20
          dead_peers <<- n                        -- 11
21
        
22
    MyOp.identity(res!19)                         -- 5
23
    TSOracle oracle_service_id
24
    consensus <- TSOracle.ts_frequency(res, tolerance, threshold, err_value)                             -- 6
25
  <- consensus, dead_peers                        -- 12
Copied!
That script is probably a little more involved than what you've seen so far. So let's work through the script: In order to get out set of timestamps, we determine the Kademlia neighbors (1) and then proceed to request a timestamp from each of those peers (2) in parallel (3). In an ideal world, each peers responds with a timestamp and the stream variable res (4) fills up with the 20 values from the twenty neighbors, which we then fold (5) and push to our consensus service (6). Alas, life in distributed isn't quite that simple since there are no guarantees that a peer is actually available to connect or provide a service response. Since we may never actually connect to a peer (7), we can't expect an error response meaning that we get a silent fail at (2) and no write to the stream res. Subsequently, this leads to the fail of the fold operation (5) since fewer than the expected twenty items are in the stream and the operation (5) ends up timing out waiting for a never-to-arrive timestamp.
In order to deal with this issue, we introduce a sleep operation (8) with the builtin Peer.timeout and run that in parallel to the attempted connection for peer n (3) essentially setting up a race condition to write to the stream: if the peer (on n, 7) behaves, we write the timestamp to res(2) and make a note of that successful operation (9); else, we write a dummy value, i.e., err_value, into the stream (10) and make a note of the delinquent peer (11). Recall, we filter out the dummy err_value at the service level.
Once we get our consensus result (6), we return it as well as the array of unavailable peers (12). And that's all there is.
In order to execute our workflow, we can use Aqua's aqua run CLI without having to manually compile the script:
1
aqua run \
2
    -a /dns4/kras-04.fluence.dev/tcp/19001/wss/p2p/12D3KooWFEwNWcHqi9rtsmDhsYcDbRUCDXH84RC4FW6UfsFWaoHi \
3
    -i aqua/ts_oracle.aqua \
4
    -f 'ts_oracle_with_consensus(10, 0.66, 0, "12D3KooWFEwNWcHqi9rtsmDhsYcDbRUCDXH84RC4FW6UfsFWaoHi", "61a86f67-ffc2-4dea-8746-fd4f04d9c75b")'
Copied!
If you are new to aqua run, the CLI functionality provided by the aqua package:
aqua run --help for all your immediate needs
the -i flag denotes the location of our reference aqua file
the -a flag denotes the multi-address of our connection peer/relay
the -f flag handles the meat of the call:
specify the aqua function name
provide the parameter values matching the function signature
Upon execution of our aqua run client, we get the following result which may be drastically different for you:
1
Your peerId: 12D3KooWEAhnNDjnh7C9Jba4Yn3EPcK6FJYRMZmaEQuqDnkb9UQf
2
[
3
  {
4
    "consensus": true,
5
    "consensus_ts": 1637883531844,
6
    "err_str": "",
7
    "n": 15,
8
    "support": 14
9
  },
10
  [
11
    "12D3KooWAKNos2KogexTXhrkMZzFYpLHuWJ4PgoAhurSAv7o5CWA",
12
    "12D3KooWHCJbJKGDfCgHSoCuK9q4STyRnVveqLoXAPBbXHTZx9Cv",
13
    "12D3KooWMigkP4jkVyufq5JnDJL6nXvyjeaDNpRfEZqQhsG3sYCU",
14
    "12D3KooWDcpWuyrMTDinqNgmXAuRdfd2mTdY9VoXZSAet2pDzh6r"
15
  ]
16
]
Copied!
Recall, that the maximum number of peers pulling from a Kademlia is 20 -- the default value set by the Fluence team. As discussed above, not all nodes may be available at any given time and at the time of this writing, the following four nodes were indeed not providing a timestamp response:
1
[
2
    "12D3KooWAKNos2KogexTXhrkMZzFYpLHuWJ4PgoAhurSAv7o5CWA",
3
    "12D3KooWHCJbJKGDfCgHSoCuK9q4STyRnVveqLoXAPBbXHTZx9Cv",
4
    "12D3KooWMigkP4jkVyufq5JnDJL6nXvyjeaDNpRfEZqQhsG3sYCU",
5
    "12D3KooWDcpWuyrMTDinqNgmXAuRdfd2mTdY9VoXZSAet2pDzh6r"
6
  ]
Copied!
That leaves us with a smaller timestamp pool to run through our consensus algorithm than anticipated. Please note that it is up to the consensus algorithm design(er) to set the minimum acceptable number of inputs deemed necessary to produce a sensible and acceptable result. In our case, we run fast and loose as evident in our service implementation discussed above, and go with what we get as long as we get at least one timestamp.
With a tolerance of ten (10) milli-seconds and a consensus threshold of 2/3 (0.66), we indeed attain a consensus for the 1637883531844 value with support from 14 out of 15 timestamps:
1
   {
2
    "consensus": true,
3
    "consensus_ts": 1637883531844,
4
    "err_str": "",
5
    "n": 15,
6
    "support": 14
7
  },
Copied!
We can make adjustments to the tolerance parameter:
1
aqua run \
2
    -a /dns4/kras-04.fluence.dev/tcp/19001/wss/p2p/12D3KooWFEwNWcHqi9rtsmDhsYcDbRUCDXH84RC4FW6UfsFWaoHi \
3
    -i aqua/ts_oracle.aqua \
4
    -f 'ts_oracle_with_consensus(0, 0.66, 0, "12D3KooWFEwNWcHqi9rtsmDhsYcDbRUCDXH84RC4FW6UfsFWaoHi", "61a86f67-ffc2-4dea-8746-fd4f04d9c75b")'
Copied!
Which does not result in a consensus timestamp given the same threshold value:
1
Your peerId: 12D3KooWP7vAR462JgoagUzGA8s9YccQZ7wsuGigFof7sajiGThr
2
[
3
  {
4
    "consensus": false,
5
    "consensus_ts": 1637884102677,
6
    "err_str": "",
7
    "n": 15,
8
    "support": 0
9
  },
10
  [
11
    "12D3KooWAKNos2KogexTXhrkMZzFYpLHuWJ4PgoAhurSAv7o5CWA",
12
    "12D3KooWHCJbJKGDfCgHSoCuK9q4STyRnVveqLoXAPBbXHTZx9Cv",
13
    "12D3KooWMigkP4jkVyufq5JnDJL6nXvyjeaDNpRfEZqQhsG3sYCU",
14
    "12D3KooWDcpWuyrMTDinqNgmXAuRdfd2mTdY9VoXZSAet2pDzh6r"
15
  ]
16
]
Copied!
We encourage you to experiment and tweak the parameters both for the consensus algorithm and the timeout settings. Obviously, longer routes make for more timestamp variance even if each timestamp called is "true."
Summary
Fluence and Aqua make it easy to create and implement decentralized oracle and consensus algorithms using Fluence's off-chain peer-to-peer network and tool set.
To further your understanding of creating decentralized off-chain (compute) oracles with Fluence and Aqua, experiment with both the consensus methodology and, of course, the oracle sources. Instead of timestamps, try your hand on crypto price/pairs and associated liquidity data, election exit polls or sports scores. Enjoy!
4. Service Composition And Reuse With Aqua
Aquamarine
Last modified 27d ago
Copy link
Contents
Overview
Creating A Decentralized Timestamp Oracle
Summary