You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

8.2 KiB

Auth route for scrapers

(Find this issue with git grep YNQAQKJS)

Test curl commands

Export the scraper API's URL prefix to an environment variable:

export API=http://127.0.0.1:4000/scraper

Put your API key into a header file, like this:

X-ApiKey: bad_password

Call it "scraper-secret.txt" or something else obviously secret. Don't check it into Git. The key will expire every 30 days and need to be rotated manually. (for now)

New versions of Curl can load headers from a text file. All commands will use this feature to load the API key.

curl --header @scraper-secret.txt $API/api/test

Should return "You're valid!"

curl --header @scraper-secret.txt $API/v1/server_list

Should return a JSON object listing all the servers.

curl --header @scraper-secret.txt $API/v1/server/aliens_wildland/api/v1/dir/

Proxies into the "aliens_wildland" server and retrieves a JSON object listing the file server root. (The server must be running a new version of ptth_server which can serve the JSON API)

curl --header @scraper-secret.txt $API/v1/server/aliens_wildland/api/v1/dir/src/

Same, but retrieves the listing for "/src".

curl --header @scraper-secret.txt $API/v1/server/aliens_wildland/files/src/tests.rs

There is no special API for retrieving files yet - But the existing server API will be is proxied through the new scraper API on the relay.

curl --header @scraper-secret.txt $API/v1/server/aliens_wildland/files/src/tests.rs

PTTH supports HEAD requests. This request will yield a "204 No Content", with the "content-length" header.

curl --header @scraper-secret.txt -H "range: bytes=100-199" $API/v1/server/aliens_wildland/files/src/tests.rs

PTTH supports byte range requests. This request will skip 100 bytes into the file, and read 100 bytes.

To avoid fence-post errors, most programming languages use half-open ranges. e.g. 0..3 means "0, 1, 2". However, HTTP byte ranges are closed ranges. e.g. 0..3 means "0, 1, 2, 3". So 100-199 means 199 is the last byte retrieved.

By polling with HEAD and byte range requests, a scraper client can approximate tail -f behavior of a server-side file.

wget --continue --execute robots=off --no-parent --recursive --header "$(<scraper-secret.txt)" $API/v1/server/aliens_wildland/files/crates/

Use wget's recursive spidering to download all the files in a folder. The human-friendly HTML interface is exposed through the scraper API, so this will also download the HTML directory listings.

  • --continue uses the server's content-length header to skip over files that are already fully downloaded to local disk. Partial downloads will be resumed where they left off, which is fine for long-running log files that may append new data but not modify old data.
  • --execute robots=off disables wget's handling of robots.txt. We know we're a robot, the server doesn't care, it's fine.
  • --no-parent prevents the ../ links from accidentally causing infinite recursion.
  • --recursive causes wget to recurse into individual files, and into subdirectories.
  • --header $(<scraper-secret.txt) tells Bash to load the secret API key from disk and send it to wget. The secret will leak into the process list, but at least it won't leak into your bash_history file.

Problem statement

PTTH has 2 auth routes:

  • A fixed API key for servers
  • Whatever the end user puts in front of the HTML client

"Whatever" is hard for scrapers to deal with. This barrier to scraping is blocking these issues:

  • EOTPXGR3 Remote tail -f
  • UPAQ3ZPT Audit logging of the relay itself
  • YMFMSV2R Add Prometheus metrics

Proposal

Add a 3rd auth route meeting these criteria:

  • Enabled by a feature flag, disabled by default
  • Bootstrapped by the user-friendly HTML frontend
  • Suitable for headless automated scrapers

It will probably involve an API key like the servers use. Public-key crypto is stronger, but involves more work. I think we should plan to start with something weak, and also plan to deprecate it once something stronger is ready.

Proposed impl plan

  • (X) Add feature flags to ptth_relay.toml for dev mode and scrapers
  • (X) Make sure Docker release CAN build
  • (X) Add hash of 1 scraper key to ptth_relay.toml, with 1 week expiration
  • (X) Accept scraper key for some testing endpoint
  • (X) (POC) Test with curl
  • (X) Clean up scraper endpoint
  • (X) Add (almost) end-to-end tests for test scraper endpoint
  • (X) Thread server endpoints through relay scraper auth
  • (don't care) Add tests for other scraper endpoints
  • (don't care) Factor v1 API into v1 module
  • (X) Add real scraper endpoints
  • ( ) Manually create SQLite DB for scraper keys, add 1 hash
  • ( ) Impl DB reads
  • ( ) Remove scraper key from config file
  • ( ) Make sure cargo test passes and Docker CAN build
  • ( ) (MVP) Test with curl
  • ( ) Impl and test DB init / migration
  • ( ) Impl DB writes (Add / revoke keys) as CLI commands
  • ( ) Implement API (Behind X-Email auth) for that, test with curl
  • ( ) Set up mitmproxy or something to add X-Email header in dev env
  • ( ) Implement web UI (Behind X-Email)

POC is the proof-of-concept - At this point we will know that in theory the feature can work.

MVP is the first deployable version - I could put it in prod, manually fudge the SQLite DB to add a 1-month key, and let people start building scrapers.

Details:

Dev mode will allow anonymous users to generate scraper keys. In prod mode, (the default) clients will need to have the X-Email header set or use a scraper key to do anything.

Design the DB so that the servers can share it one day.

Design the API so that new types of auth / keys can be added one day, and the old ones deprecated.

Endpoints needed:

  • (X) Query server list
  • (X) Query directory in server
  • (not needed) GET file with byte range (identical to frontend file API)

These will all be JSON for now since Python, Rust, C++, C#, etc. can handle it. For compatibility with wget spidering, I might do XML or HTML that's machine-readable. We'll see.

DB / UI impl

Sprint 1:

  • Look up keys by their hash
  • not_before
  • not_after
  • name
  • X-Email associated with key

Sprint 2:

  • UI to generate / revoke keys

SQL schema

Migration

create table scraper_keys (
	hash text primary key,        -- Using blake3 for this because it's not a password
	not_before integer not null,  -- Seconds since epoch 
	not_after integer not null,   -- Seconds since epoch
	name text not null,           -- Human-friendly nickname
	email text not null           -- Email address that created the key
);

Look up hash

select not_before, not_after name, email 
from scraper_keys 
where 
	hash = $1 and 
	strftime ('%s') >= not_before and 
	strftime ('%s') < not_after
;

Create key

-- Generate entropy in app code
insert into scraper_keys (
	hash,
	not_before,
	not_after,
	name,
	email
) values (
	$1, 
	strftime ('%s'),
	strftime ('%s') + 2592000,
	$4,
	$5
);

-- Respond to client with plaintext key and then forget it.
-- If a network blip causes the key to evaporate, the client should revoke it.

Revoke key


Decision journal

Who generates the API key? The scraper client, or the PTTH relay server?

The precedent from big cloud vendors seems to be that the server generates tokens. This is probably to avoid a situation where clients with vulnerable crypto code or just bad code generate low-entropy keys. By putting that responsibility on the server, the server can enforce high-entropy keys.

Should the key rotate? If so, how?

The key should at least expire. If it expires every 30 or 90 days, then a human is slightly inconvenienced to service their scraper regularly.

When adding other features, we must consider the use cases:

  1. A really dumb Bash script that shells out to curl
  2. A Python script
  3. A sophisticated desktop app in C#, Rust, or C++
  4. Eventually replacing the fixed API keys used in ptth_server

For the Bash script, rotation will probably be difficult, and I'm okay if our support for that is merely "It'll work for 30 days at a time, then you need to rotate keys manually."

For the Python script, rotation could be automated, but cryptography is still probably difficult. I think some AWS services require actual crypto keys, and not just high-entropy password keys.

For the sophisticated desktop app, cryptography is on the table, but this is the least likely use case to ever happen, too.