ETA Prediction System With Kafka & DynamoDB

Not too long ago we have seen a giant enhance in the usage of on-demand logistics providers, comparable to on-line procuring and meals supply.

Most of those knowledge functions present a close to real-time monitoring of the ETA when you place the order. Constructing a scalable, distributed, and real-time ETA prediction system is a tricky activity, however what if we might simplify its design? We’ll break our system into items such that every element is liable for one main job.

Let’s check out elements that represent the system.

1. Supply driver/rider app – The Android/iOS app put in on a supply individual’s machine.
2. Buyer app – The Android/iOS app put in on a buyer’s machine.
3. Rockset – The question engine powering all of the fashions and providers.
4. Message queue – Used for transferring knowledge between numerous elements. For this instance, we’ll use Kafka.
5. Key-value storage – Used for storing orders and parameters for the mannequin. For this instance, we’ll use DynamoDB.

Inputs to the Mannequin

Driver Location

To get an correct ETA estimation, you will want the supply individual’s place, particularly the latitude and longitude. You will get this info simply by way of GPS in a tool. A name to the machine GPS supplier returns latitude, longitude, and the accuracy of the situation in meters.

You’ll be able to run a background service within the app that retrieves the GPS coordinates each 10 seconds. The coordinates, as such, are too fine-grained to make a prediction. To extend the granularity of the GPS, we will likely be utilizing the idea of geohash. A geohash is a standardized N-letter hash of a location that represents an space of M sq. miles. N and M are inversely proportional, so a bigger N represents a smaller space M. You’ll be able to seek advice from this for more information on geohash.

There are tons of libraries out there to transform latitude-longitude to geohash. Right here we’ll be utilizing geo by davidmoten to get a 6-7 letter geohash.

The service then pushes the geohash together with the coordinates to a Kafka matter. Rockset ingests knowledge from this Kafka matter and updates it into a group known as areas.

Orders

The orders positioned by a buyer are saved in DynamoDB for additional processing. An order typically goes via a life cycle consisting of the next states:

• CREATED
• PROCESSING
• CONFIRMED
• CANCELED
• IN TRANSIT
• DELIVERED

The entire above state modifications are up to date in DynamoDB together with extra knowledge such because the supply location, vacation spot location, order particulars, and many others. As soon as an order is delivered, the precise time of arrival can also be saved within the database.

Rockset additionally ingests updates from DynamoDB orders desk and updates it into a group known as orders.

ML Mannequin

Exponential Smoothing

We now have the precise time of arrival together with the supply and the vacation spot for order out there from the orders desk. We’ll seek advice from it as TA. You’ll be able to take the imply of all of the TA with supply as supply individual’s newest location and vacation spot as buyer’s location, and you may get an approximate ETA. Nonetheless, this isn’t that correct because it would not account for altering elements, comparable to new development actions within the space or new shorter routes to the vacation spot.

To do this, we’d like a prediction mannequin that’s simplistic and straightforward to debug and has good accuracy.

That is the place exponential smoothing comes into play. An exponentially smoothened worth is calculated utilizing the components:

S_t = Alpha * X_t + (1 – Alpha) * S_t-1

the place

• S_t => Smoothened worth at time t
• X_t => Precise worth at time t
• Alpha => Smoothing issue

In our context, S_t represents the ETA and X_t represents the newest precise time of arrival for a source-destination pair in our orders desk.

ETA_t = Alpha * TA_t + (1 – Alpha) * ETA_t-1

Rockset

The serving layer for the present system must fulfill three main standards:

1. Capacity to deal with hundreds of thousands of writes per minute – Every supply individual’s app will likely be pushing GPS coordinates each 5-10 seconds, which can result in a brand new ETA. A typical giant scale meals supply firm has virtually 100K supply individuals.
2. The info fetch latency must be minimal – For an amazing UX, we must always have the ability to replace ETA on the shopper app as quickly as it’s up to date.
3. Capacity to deal with schema modifications on the fly – we are able to retailer extra metadata comparable to ETA prediction accuracy and mannequin model sooner or later. We do not wish to create a brand new knowledge supply each time we add a brand new subject.

Rockset satisfies all of them. It has:

1. Dynamic Scaling – Extra sources are added as and when wanted to deal with giant volumes of knowledge.
2. Distributed Question Processing – Parallelisation of queries throughout a number of nodes to reduce latency
3. Schemaless Ingest – to help schema modifications on the fly.

Rockset has a built-in connector to Apache Kafka. We are able to use this Kafka connector to ingest location knowledge of the supply individual.

To carry out exponential smoothing in Rockset, we create two Question Lambdas. Question Lambdas in Rockset are named, parameterized SQL queries saved in Rockset that may be executed from a devoted REST endpoint.

1. calculate_ETA: The Question Lambda expects alpha, supply, and vacation spot as a parameter. It returns an exponentially smoothened ETA. It runs the next question to get the specified outcome:

SELECT
    (:alpha * SUM(time period)) + (POW((1 - :alpha), MAX(idx))* MIN_BY(ta_i, time_i)) as ans
FROM
    (
        (
            SELECT
                order_id,
                ta_i,
                (ta_i * POW((1 - :alpha), (idx - 1))) AS time period,
                time_i,
                idx
            FROM
                (
                    SELECT
                        order_id,
                        CAST(ta AS int) as ta_i,
                        time_i,
                        ROW_NUMBER() OVER(
                            ORDER BY
                                time_i DESC, order_id ASC
                        ) AS idx
                    FROM
                        commons.orders_fixed
                    WHERE
                        source_geohash = :supply
                        AND
                        destination_geohash = :vacation spot
                    ORDER BY
                        time_i DESC, order_id ASC
                ) AS idx
        ) AS phrases
    )

2. calculate_speed: This Question Lambda requires order_id as param and returns the common pace of the supply individual whereas in transit. It runs the next question:

SELECT
    SUM(ST_DISTANCE(prev_geo, geo) /(ts - prev_ts)) / COUNT(*) AS pace
FROM
    (
        SELECT
            geo,
            LEAD(geo, 1) OVER(
                ORDER BY
                    ts DESC
            ) AS prev_geo,
            ts,
            LEAD(ts, 1) OVER(
                ORDER BY
                    ts DESC
            ) AS prev_ts
        FROM
            (
                SELECT
                    ST_GEOGPOINT(CAST(lng AS double), CAST(lat AS double)) AS geo,
                    order_id,
                    CAST(timestamp as int) AS ts
                FROM
                    commons.areas
                WHERE
                    order_id = :order_id
            ) AS ts
    ) As pace

Predict the ETA

The shopper app initiates the request to foretell the ETA. It passes the order id within the API name.

The request goes to the question service. Question service performs the next capabilities:

1. Fetch the newest smoothing elements Alpha and Beta from DynamoDB. Right here, Alpha is the smoothing parameter and Beta is the load assigned to historic ETA whereas calculating the ultimate ETA. Refer step 6 for extra particulars
2. Fetch the vacation spot geohash for the order id.
3. Fetch the present driver geohash from the areas assortment.
4. Set off calculate_ETA Question Lamba in Rockset with smoothing issue alpha as param and driver geohash as supply geohash and vacation spot geohash from step 2. Let’s name this historic ETA.

curl --request POST 
--url https://api.rs2.usw2.rockset.com/v1/orgs/self/ws/commons/lambdas/calculateETA/variations/f7d73fb5a786076c 
-H 'Authorization: YOUR ROCKSET API KEY' 
-H 'Content material-Sort: software/json' 
  -d '{
    "parameters": [
      {
        "name": "alpha",
        "type": "float",
        "value": "0.7"
      },
      {
        "name": "destination",
        "type": "string",
        "value": "tdr38d"
      },
      {
        "name": "source",
        "type": "string",
        "value": "tdr706"
      }
    ]
  }'

5. Set off calculate_speed Question Lambda in Rockset with present order id as param

curl --request POST 
--url https://api.rs2.usw2.rockset.com/v1/orgs/self/ws/commons/lambdas/calculate_speed/variations/cadaf89cba111c06 
-H 'Authorization: YOUR ROCKSET API KEY' 
-H 'Content material-Sort: software/json' 
  -d '{
    "parameters": [
      {
        "name": "order_id",
        "type": "string",
        "value": "abc"
      }
    ]
  }'

6. The expected ETA is then calculated by question service as

Predicted ETA = Beta * (historic ETA) + (1 – Beta) * distance(driver, vacation spot)/pace

The expected ETA is then returned to the shopper app.

Suggestions Loop

ML fashions require retraining in order that their predictions are correct. In our situation, it’s fairly essential to re-train the ML mannequin in order to account for altering climate circumstances, festivals, and many others. That is the place the parameter tuning service comes into play.

Parameter Tuning Service

As soon as an ETA is predicted, you possibly can retailer the expected ETA, and the precise ETA in a group known as predictions. The first motivation to retailer this knowledge in Rockset as a substitute of another datastore is to create a real-time dashboard for measuring the accuracy of the mannequin. That is wanted to verify the shoppers don’t see absurd ETA values of their apps.

The subsequent query is decide the smoothing issue Alpha. To resolve this difficulty, we create a parameter tuning service, which is only a Flink batch Job. We fetch all of the historic ETAs and TAs for orders for the previous 7-30 days. We use the distinction in these ETAs to calculate acceptable Alpha and Beta values. This may be performed utilizing a easy mannequin comparable to logistic regression.

As soon as the service calculates the Alpha and Beta parameters, they’re saved in DynamoDB in a desk named smoothing_parameters. The question service fetches the parameters from this desk when it receives a request from the patron app.

You’ll be able to practice the parameter tuning mannequin as soon as per week utilizing ETA knowledge in areas assortment.

Conclusion

The structure is designed to deal with greater than 1,000,000 requests per minute whereas being versatile sufficient to help the scaling of the applying on the fly. The structure additionally permits builders to modify or insert elements comparable to including new options (e.g. climate) or including a filter layer to refine the ETA predictions. Right here, Rockset helps us clear up three main necessities:

1. Low-latency advanced queries – Rockset permits us to make difficult queries comparable to exponential smoothing with simply an API name. That is performed by leveraging Question Lambdas. The Lambdas additionally help parameters that enable us to question for various areas.
2. Extremely scalable real-time ingestion – In case you have roughly 100K drivers in your platform and every of their apps sends a GPS location each 5 seconds, then you might be coping with a throughput of 1.2 million requests per minute. Rockset permits us to question this knowledge inside seconds of occasions occurring.
3. Knowledge from a number of sources – Rockset permits us to ingest from a number of sources, comparable to Kafka and DynamoDB, utilizing absolutely managed connectors that require minimal configuration.

Kartik Khare has been a Knowledge Engineer for 4 years and has additionally been running a blog about deep-dives on Massive Knowledge Programs on a private weblog and Medium. He at the moment works at Walmart Labs the place he works on the Realtime ML platforms. Previous to that, he was working for OlaCabs the place he was concerned in designing realtime surge pricing and advice programs.