Hark lets you quickly build serverless data pipelines without managing any infrastructure.
Hark is for you if:
- You use AWS.
- You use Python for data engineering or business process pipelines.
- You don't want to manage a task platform (Airflow, Celery, etc).
Key features:
- First-class local testing (there's a local Hark runtime).
- Concurrency primitives for multi-threaded pipelines.
- Zero infrastructure management and minimal maintenance.
Quick start: Build an AWS Lambda pipeline in 2 minutes.
Comparisons:
- Like Apache Airflow, but without infrastructure to manage.
- Like AWS Step Functions but cloud-portable and locally testable.
- Like Serverless Framework, but handles runtime glue logic in addition to deployment.
Status: Hark works well for small workflows: 5-10 Lambda invocations. Larger workflows may cause problems, and there is a known issue caused by DynamoDB restrictions (#12).
Hark was Presented at PyCon Africa 2020. Watch the presentation, or check out the demos.
Hark is growing rapidly, and contributions are welcome.
Hark is for you if:
- Your data is in AWS
- You use Python for processing data, or writing business process workflows.
- You don't want to deploy and manage a task platform (Airflow, Celery, etc).
Data in: You can invoke Hark like any Lambda function (AWS cli, S3 trigger, API gateway, etc).
Data out: Use the Python libraries you already have for database access. Hark just connects them together.
Development: Hark runs locally, so you can thoroughly test Hark programs before deployment (using minio and localstack for any additional infrastructure that your code uses.
Operating: Hark enables contextual cross-thread logging and stacktraces out of the box, since the entire application is described in one place.
Hark is like... | But... |
---|---|
AWS Step Functions | Hark programs aren't bound to AWS and don't use Step Functions under the hood (just plain Lambda + DynamoDB). |
Orchestrators (Apache Airflow, etc) | You don't have to manage infrastructure, or think in terms of DAGs, and you can test everything locally. |
Task runners (Celery, etc) | You don't have to manage infrastructure. |
Azure Durable Functions | While powerful, Durable Functions (subjectively) feel complex - their behaviour isn't always obvious. |
All you need:
- An AWS account, and AWS CLI configured.
- A Python 3.8 virtual environment
Hark is built with Python, and distributed as a Python package. To install it, run in a new virtualenv:
pip install hark-lang
This gives you the hark
executable. Try hark -h
.
Initialise the project with a few template files:
hark init
Copy the following snippet into service.hk
:
// Import the processing functions defined in Python
import(process_video_step1, src.video, 2);
import(process_video_step2, src.video, 3);
import(process_video_step3, src.video, 3);
import(process_video_final_step, src.video, 3);
// Process a named file
fn process_csv(key) {
a = async process_video_step1(bucket, key);
b = async process_video_step2(bucket, key, await a);
c = async process_video_step3(bucket, key, await b);
process_video_final_step(bucket, key, await c);
}
Run it locally to test:
hark service.hk -f on_upload filename.csv
And deploy the service to your AWS account (requires AWS credentials and
AWS_DEFAULT_REGION
to be defined):
hark deploy
Read more about what this actually creates.
Finally, invoke it in AWS (-f main
is optional, as before):
hark invoke -f main your_bucket filename.csv
Read more...
This is useful when a set computations are related, and must be kept together.
/**
* Return f(x) + g(x), computing f(x) and g(x) in parallel in two separate
* threads (Lambda invocations in AWS).
*/
fn compute(x) {
a = async f(x); // Start computing f(x) in a new thread
b = async g(x); // Likewise with g(x)
await a + await b; // Stop this thread, and resume when {a, b} are ready
}
Traditional approach: Manually store intermediate results in an external
database, and build the synchronisation logic into the cloud functions f
and
g
, or use an orchestrator service.
Use this approach when each individual function may take several minutes (and hence, together would break the 5 minute AWS Lambda limit).
/**
* Compute f(g(h(x))), using a separate lambda invocation for each
* function call.
*/
fn pipeline(x) {
a = async h(x);
b = async g(await a);
f(await b);
}
Traditional approach: This is functionally similar to a "chain" of AWS Lambda functions and SQS queues.
Hark functions are first-class, and can be passed around (closures and anonymous functions are planned, giving Hark object-oriented capabilities).
/**
* Compute [f(element) for element in x], using a separate lambda invocation for
* each application of f.
*/
fn map(f, x, accumulator) {
if nullp(x) {
accumulator
}
else {
// The Hark compiler has tail-recursion optimisation
map(func, rest(x), append(accumulator, async f(first(x))))
}
}
This could be used like:
fn add2(x) {
x + 2
}
fn main() {
futures = map(add2, [1, 2, 3, 4], []);
// ...
}
The syntax should look familiar, but there are a couple of things to point out.
Every expression must return a value, so there is no return
statement. The
last expression in a 'block' (expressions between {
and }
) is returned
implicitly.
fn foo() {
"something"
}
fn main() {
print(foo()) // -> prints "something"
}
... when there is more than one expression in a block.
This is ok:
fn main() {
print("done")
}
So is this:
fn main() {
print("one");
print("two")
}
And this:
fn main() {
print("one");
print("two");
}
But this is not ok:
fn main() {
print("one") // <- missing semicolon!
print("two")
}
In this snippet, "Hello Worlds!" is actually printed twice. First in bar
, then
in main
.
fn bar() {
print("Hello Worlds!")
}
fn main() {
print(bar())
}
$> hark -q service.hk
Hello Worlds!
Hello Worlds!
Think about it like this: An if
expression represents a choice between
values.
v = if something { true_value } else { false_value };
// if 'something' is not true, v is set to null
v = if something { value };
Why is this not a library/DSL in Python?
When Hark threads wait on a Future, they stop completely. The Lambda function saves the machine state and then terminates. When the Future resolves, the resolving thread restarts any waiting threads by invoking new Lambdas to pick up execution.
To achieve the same thing in Python, the framework would need to dump the entire Python VM state to disk, and then reload it at a later point -- this may be possible, but would certainly be non-trivial. An alternative approach would be to build a langauge on top of Python that looked similar to Python, but hark wrong because it was really faking things under the hood.
How is Hark like Go?
Goroutines are very lightweight, while Hark async
functions are pretty heavy --
they involve creating a new Lambda (or process, when running locally).
Hark's concurrency model is similar to Go's, but channels are not fully implemented so data can only be sent to/from a thread at call/return points.
Is this an infrastructure-as-code tool?
No, Hark does not do general-purpose infrastructure management. There are already great tools to do that (Terraform, Pulumi, Serverless Framework, etc).
Instead, Hark reduces the amount of infrastructure you need. Instead of a distinct Lambda function for every piece of application logic, you only need the core Hark interpreter (purely serverless) infrastructure.
Hark will happily manage that infrastructure for you (through hark deploy
and
hark destroy
), or you can set it up with your in-house custom system.
Hark is beta quality, which means that it's not thoroughly tested or feature complete. This is a non-exhaustive list.
Only one Hark program file is supported, but a module/package system is planned.
There's no error handling - if your function fails, you'll have to restart the whole process manually. An exception handling system is planned.
Function inputs and outputs aren't typed. This is a limitation, and will be fixed soon, probably using ProtoBufs as the interface definition language.
Currently you can only call Hark or Python functions -- arbitrary microservices
can't be called. Before Hark v1.0 is released, this will be possible. You will
be able to call a long-running third party service (e.g. an AWS ML service) as a
normal Hark function and await
on the result.
Hark started because we couldn't find any data engineering tools that were productive and feel like software engineering. As an industry, we've spent decades growing a wealth of computer science knowledge, but building data pipelines in $IaC, or manually crafting workflow DAGs with $AutomationTool, feels more like hardware than software.
Apache License (Version 2.0). See LICENSE for details.
The end. Here's a spaceship. Hacks and glory await.
`. ___
__,' __`. _..----....____
__...--.'``;. ,. ;``--..__ .' ,-._ _.-'
_..-''-------' `' `' `' O ``-''._ (,;') _,'
,'________________ \`-._`-','
`._ ```````````------...___ '-.._'-:
```--.._ ,. ````--...__\-.
`.--. `-` ____ | |`
`. `. ,'`````. ; ;`
`._`. __________ `. \'__/`
`-:._____/______/___/____`. \ `
| `._ `. \
`._________`-. `. `.___
SSt `------'`