zchpc 0.1.0

Client for the ZCHPC platform: GPUs, managed databases, caches and media buckets.

zchpc

The Python client for ZCHPC: rent GPU slices, and provision managed databases, caches and S3-compatible media buckets.

pip install zchpc

Nothing is required to talk to the API itself. The drivers are pulled in only when you actually connect to a resource, so a bucket user never installs a Postgres driver:

pip install "zchpc[postgres]"   # psycopg
pip install "zchpc[cache]"      # redis
pip install "zchpc[buckets]"    # boto3
pip install "zchpc[gpu]"        # paramiko
pip install "zchpc[all]"        # all of them

Rent a GPU

with destroys the rental on the way out, including when the code in the middle raises. A GPU bills by the second, and the usual way a box gets left running is an exception between renting it and destroying it.

from zchpc import ZCHPC

z = ZCHPC("gpv_YOUR_API_KEY")        # or set ZCHPC_API_KEY

with z.gpus.rent(gpu_mem_gb=8, hours=1).wait() as gpu:
    print(gpu.gpu_info())            # Tesla V100S-PCIE-32GB, 32768 MiB, 0 MiB
    print(gpu.run("nvidia-smi"))

hours is prepaid, so ask for what you need: the API default of 24 is a full day of rent.

Databases

db = z.databases.first_running("postgres") or z.databases.create().wait()

db.sql("CREATE TABLE IF NOT EXISTS notes (id serial PRIMARY KEY, body text)")
db.executemany("INSERT INTO notes (body) VALUES (%s)", [("first",), ("second",)])

print(db.sql("SELECT id, body FROM notes ORDER BY id"))
print(db.tables())
print(db.uri)                        # the raw connection string

The client is a convenience, not a wall. db.uri is a standard connection string and db.connect() is a real driver connection, so you can drop to plain SQL whenever the wrapper is in the way.

Cache

cache = z.cache.first_running() or z.cache.create().wait()

cache.set("session:42", "abc", ttl=60)
print(cache.ttl("session:42"))       # 60  (-1 = no expiry, -2 = no key)
cache.incr("hits")                   # atomic, safe across processes
cache.flush_prefix("session:")       # scoped delete, never FLUSHALL

Buckets

bucket = z.buckets.first() or z.buckets.create("media")

bucket.put("photos/cat.jpg", open("cat.jpg", "rb").read(),
           content_type="image/jpeg")

print(bucket.presign("photos/cat.jpg", expires=300))   # expiring share link
print(bucket.presign_put("uploads/from-browser.png"))  # browser-direct upload

Path-style addressing is set for you. Amazon puts the bucket in the hostname and every SDK defaults to that, but this endpoint serves buckets as a path, so doing it by hand is the usual first thing to go wrong.

Errors

Typed by what you can do about them: AuthError, InsufficientCredit, LimitReached, NotFoundError, ProvisioningError, WaitTimeout, MissingDependency.

License

MIT