Administration#
Resource Requirements#
conda-store-serveris simply a web server and should not require any specific resources. 1 GB of RAM and 1 CPU should be plenty.conda-store-workerdoes the actual builds of the Conda environments. Solving for Conda environments can take a lot of memory in some circumstances. So make sure to allocate at least 4 GB of RAM to the worker along with at least one CPU.
Performance#
There are several parts of conda-store to consider for performance. We have tried to list them in order of performance impact that may be seen.
Worker Storage#
When conda-store builds a given environment it has to locally install
the environment in the directory specified in the
Traitlets
configuration CondaStore.store_directroy. Conda environments consist
of many hardlinks to small files. This means that the
store_directory is limited to the number of
IOPS the directory can
perform. Many cloud providers have high performance storage
options. These include:
If you do not need to mount the environments via NFS into the containers we highly recommend not using NFS and using traditional block storage. Not only is it significantly cheaper but the IOPs performance will be better as well.
If you want to mount the environments in containers or running VMs NFS may be a good option for you. With NFS many cloud providers provide a high performance filesystem option at a significant premium in cost. Example of these include GCP Filestore, AWS EFS, and Azure files. Choosing an nfs storage option with low IOPS will result in long environment install times.
Network Speed#
While Conda does its best to cache packages, it will have to reach out
to download the repodata.json along with the packages as well. Thus
network speeds may be important. Typically cloud environments have
plenty fast Internet.
S3 Storage#
All build artifacts from conda-store are stored in object storage that behaves S3 like. S3 traditionally has great performance if you use the cloud provider implementation.
Configuration#
conda-store is configured via Traitlets. Originally this configuration was done via command line options but as the options grew this seems untenable. conda-store server and worker can be launched via configuration easily.
conda-store-server --config <path-to-conda-store-config.py>
conda-store-worker --config <path-to-conda-store-config.py>
Below we outline the options for conda-store.
conda_store_server.app.CondaStore#
CondaStore.storage_class configures the storage backend to use for
storing build artifacts from
conda-store. S3 storage is
the default. File based storage is also supported but not nearly as
well tested.
CondaStore.conda_solve_platforms configures which platforms to solve environments for, via conda-lock. It must include the current platform conda-store is running on. By default, contains only the platform on which conda-store is running.
CondaStore.store_directory is the directory used for conda-store to
build the environments.
CondaStore.build_directory template used to form the directory for
storing Conda environment builds. Available keys: store_directory,
namespace, name. The default will put all built environments in the
same namespace within the same directory.
CondaStore.environment_directory template used to form the directory
for symlinking Conda environment builds. Available keys:
store_directory, namespace, name. The default will put all
environments in the same namespace within the same directory.
CondaStore.validate_specification callable function taking
conda_store and specification as input arguments to apply for
validating and modifying a given specification. If there are
validation issues with the environment ValueError with message will be
raised.
CondaStore.validate_action callable function taking conda_store,
namespace, and action. If there are issues with performing the given
action raise a CondaStoreError should be raised.
CondaStore.conda_command is the command to use for creation of
Conda environments. Currently mamba is the default which will
usually result in lower peak memory usage and faster builds.
CondaStore.conda_channel_alias is the url to prepend to all
shorthand Conda channels that do not specify a url. The default is
https://conda.anaconda.org.
CondaStore.conda_platforms are the platforms to download package
repodata.json from. By default includes current architecture and
noarch.
CondaStore.conda_default_channels is a list of Conda channels that
are by default added if channels within the specification is empty.
CondaStore.conda_allowed_channels is a list of Conda channels that
are allowed. This also tells conda-store which channels to prefetch
the channel repodata and channeldata from. The default is main
and conda-forge. If conda_allowed_channels is an empty list all
Channels are accepted by users.
CondaStore.conda_indexed_channels tells conda-store which channels to prefetch
the channel repodata and channeldata from. The default is main
and conda-forge.
CondaStore.conda_default_packages is a list of Conda packages that
are included by default if none are specified within the specification
dependencies.
CondaStore.conda_required_packages is a list of Conda packages that
are required upon validation of the specification dependencies. This
will not auto add the packages but instead throw an error that they
are missing.
CondaStore.conda_included_packages is a list of Conda packages that
if not specified within the specification dependencies will be auto
added.
CondaStore.pypi_default_packages is a list of PyPi packages that
are included by default if none are specified within the specification
dependencies.
CondaStore.pypi_required_packages is a list of PyPi packages that
are required upon validation of the specification dependencies. This
will not auto add the packages but instead throw an error that they
are missing.
CondaStore.pypi_included_packages is a list of PyPi packages that
if not specified within the specification dependencies will be auto
added.
CondaStore.storage_thresold storage threshold in bytes of minimum
available storage required in order to perform builds.
CondaStore.database_url is the url string for connecting to the
database. Behind the scenes SQLAlchemy
is used for the connection so consult their
docs for
connecting to your specific database. conda-store will automatically
create the tables if they do not already exist.
CondaStore.redis_url is an optional argument to a running Redis
instance. This was removed as a dependency as of release 0.4.10 due
to the need to have a simple deployment option for conda-store. See
documentation
for proper specification. This url is used by default for the Celery
broker and results backend.
CondaStore.celery_broker_url is the broker use to use for
celery. Celery supports a wide range of
brokers
each with different guarantees. By default the Redis based broker is
used if a CondaStore.redis_url if provided otherwise defaults to
sqlalchemy. It is production ready and has worked well in
practice. The url must be provided in a format that celery
understands. The default value is CondaStore.redis_url.
CondaStore.build_artifacts is the list of artifacts for conda-store
to build. By default it is all the artifacts that conda-store is
capable of building. These are the
lockfile,
YAML,
Conda-Pack, and
Docker. Currently
the lockfile one is ignored since it is always created upon build.
CondaStore.build_artifacts_kept_on_deletion is a list of artifacts
to keep after a given build is deleted. Often an administrator will
want to keep around the logs etc. of a build and the Conda solve for
the given build.
CondaStore.celery_results_backend is the backend to use for storing
all results from celery task execution. conda-store currently does not
leverage the backend results but it may be needed for future work
using celery. The backend defaults to using the Redis backend if
CondaStore.redis_url is specified otherwise uses the
CondaStore.database_url. This choice works great in
production. Please consult the celery docs on
backend.
CondaStore.default_namespace is the default namespace for
conda-store to use. All environments are built behind a given
namespace.
CondaStore.filesystem_namespace is the namespace to use for
environments picked up via CondaStoreWorker.watch_paths on the
filesystem.
CondaStore.default_uid is the uid (user id) to assign to all
files and directories in a given built environment. This setting is
useful if you want to protect environments from modification from
certain users and groups.
CondaStore.default_gid is the gid (group id) to assign to all
files and directories in a given built environment. This setting is
useful if you want to protect environments from modification from
certain users and groups.
CondaStore.default_permissions is the filesystem permissions to
assign to all files and directories in a given built environment. This
setting is useful if you want to protect environments from
modification from certain users and groups.
CondaStore.default_docker_base_image default base image used for the
Dockerized environments. Make sure to have a proper glibc within image
(highly discourage alpine/musl based images). Can also be callable
function which takes the orm.Build object as input which has access
to all attributes about the build such as installed packages, requested
packages, name, namespace, etc. This package at a minimum should have
the following packages
installed. Often
times for non-graphic and non-gpu environments glibc is enough. Hence
the default docker image library/debian:sid-slim.
CondaStore.serialize_builds DEPRECATED no longer has any effect
CondaStore.post_update_environment_build_hook is an optional configurable to
allow for custom behavior that will run after an environment’s current build changes.
conda_store_server.storage.S3Storage#
conda-store uses minio-py as a client to connect to S3 “like” object stores.
S3Storage.internal_endpoint is the internal endpoint for conda-store
reaching out to s3 bucket. This is the url that conda-store use for
get/set s3 blobs. For AWS S3 use the endpoint s3.amazonaws.com.
S3Storage.external_endpoint is the external s3 endpoint for users to
reach out to in the presigned url. This is the url that users use
for fetching s3 blobs. For AWS S3 use the endpoint s3.amazonaws.com.
S3Storage.access_key is the access key for S3 bucket.
S3Storage.secret_key is the secret key for S3 bucket.
S3Storage.region is the region to use for connecting to the S3
bucket. The default is us-east-1.
S3Storage.bucket_name is the bucket name to use for connecting to
the S3 bucket.
S3Storage.internal_secure Boolean to indicate if connecting via
http (False) or https (True) internally. The internal connection
is the url that will be exclusively used by conda-store and not shared
with users.
S3Storage.external_secure Boolean to indicate if connecting via
http (False) or https (True) internally. The external connection
is the url that will be served to users of conda-store.
S3Storage.credentials provider to use to get credentials for s3
access. see examples
https://github.com/minio/minio-py/tree/master/examples and
documentation
https://github.com/minio/minio-py/blob/master/docs/API.md#1-constructor. An
example of this could be to use minio.credentials.IamAwsProvider to
get S3 credentials via IAM.
S3Storage.credentials_args arguments to pass for creation of
credentials class.
S3Storage.credentials_kwargs keyword arguments to pass for creation
of credentials class.
conda_store_server.storage.LocalStorage#
LocalStorage.storage_path is the base directory to use for storing
build artifacts.
LocalStorage.storage_url is the base url for serving of build
artifacts. This url assumes that the base will be a static server
serving LocalStorage.storage_path.
conda_store_server.server.auth.AuthenticationBackend#
AuthenticationBackend.secret is the symmetric secret to use for
encrypting tokens.
AuthenticationBackend.jwt_algorithm is the algorithm for encrypting
the JSON Web Tokens.
AuthenticationBackend.predefined_tokens is a set of tokens with
predefined permission. This is useful for setting up service accounts
in a similar manner to how things are done with jupyterhub. Format for
the values is a dictionary with keys being the tokens and values being
the schema.AuthenticaitonToken all fields are optional.
conda_store_server.server.auth.Authentication#
Authentication.cookie_name is the name for the browser cookie used
to authenticate users.
Authentication.cookie_domain use when wanting to set a subdomain wide cookie. For example setting this to example.com would allow the cookie to be valid for example.com along with *.example.com.
Authentication.authentication_backend is the class to use for
authentication logic. The default is AuthenticationBackend and will
likely not need to change.
Authentication.authorization_backend is the class to use for
authentication logic. The default is AuthorizationBackend and will
likely not need to change.
Authentication.login_html is the HTML to display for a given user as
the login form.
conda_store_server.server.auth.DummyAuthentication#
Has all the configuration settings of Authetication. This class is
modeled after the JupyterHub DummyAuthentication
class.
DummyAuthentication.password sets a global password for all users to
login with. Effectively a static password. This rarely if ever should
be used outside of testing.
conda_store_server.server.auth.GenericOAuthAuthentication#
A provider-agnostic OAuth authentication provider. Configure
endpoints, secrets and other parameters to enable any OAuth-compatible
platform. This class is modeled after the OAuthenticator OAuth2
classes. All
configuration settings of Authentication are available.
GenericOAuthAuthentication.access_token_url is the URL used to
request an access token once app has been authorized.
GenericOAuthAuthentication.authorizie_url is the URL used to request
authorization to OAuth provider.
GenericOAuthAuthentication.client_id is the unique string that
identifies the app against the OAuth provider.
GenericOAuthAuthentication.client_secret is the secret string used
to authenticate the app against the OAuth provider.
GenericOauthAuthentication.access_scope is the permissions that will
be requested to OAuth2 provider. This is space separated list of
scopes. Generally only one scope is requested.
GenericOauthAuthentication.user_data_url is the API endpoint for
OAuth provider that returns a JSON dict with user data after the user
has successfully authenticated.
GenericOauthAuthentication.user_data_key is the key in the payload
returned by user_data_url endpoint that provides the username.
GenericOAuthAuthentication.oauth_callback_url custom callback url
especially useful when web service is behind a proxy.
GenericOAuthAuthentication.tls_verify to optionally turn of TLS
verification useful for custom signed certificates.
conda_store_server.server.auth.GithubOAuthAuthentication#
Inherits from Authentication and GenericOAuthAuthentication so
should be fully configurable from those options.
GithubOAuthAuthentication.github_url is the url for GitHub. Default
is https://github.com.
GithubOAuthAuthentication.github_api is the REST API url for
GitHub. Default is https://api.github.com.
conda_store_server.server.auth.JupyterHubOAuthAuthentication#
Inherits from Authentication and GenericOAuthAuthentication so
should be fully configurable from those options.
GithubOAuthAuthentication.jupyterhub_url is the url for connecting
to JupyterHub. The URL should not include the /hub/.
conda_store_server.server.app.CondaStoreServer#
CondaStoreServer.log_level is the level for all server
logging. Default is INFO. Common options are DEBUG, INFO,
WARNING, and ERROR.
CondaStoreServer.enable_ui a Boolean on whether to expose the UI
endpoints. Default True.
CondaStoreServer.enable_api a Boolean on whether to expose the API
endpoints. Default True.
CondaStoreServer.enable_registry a Boolean on whether to expose the
registry endpoints. Default True.
CondaStoreServer.enable_metrics a Boolean on whether to expose the
metrics endpoints. Default True.
CondaStoreServer.address is the address for the server to bind
to. The default is all IP addresses 0.0.0.0.
CondaStoreServer.port is the port for conda-store server to
use. Default is 5000.
CondaStoreServer.registry_external_url is the external hostname and
port to access docker registry cannot contain http:// or https://.
CondaStoreServer.url_prefix is the prefix URL (subdirectory) for the
entire application. All but the registry routes obey this. This is due
to the docker registry API specification not supporting url prefixes.
CondaStoreServer.authentication_class is the authentication class
for the web server to use. Default is DummyAuthentication.
CondaStoreServer.secret_key is a secret key needed for some
authentication methods, session storage, etc. TODO: remove at some
point since also used in AuthenticationBackend.
CondaStoreServer.max_page_size is maximum number of items to return
in a single UI page or API response.
CondaStoreServer.behind_proxy indicates if server is behind web
reverse proxy such as Nginx, Traefik, Apache. Will use
X-Forward-... headers to determine scheme. Do not set to true if not
behind proxy since Flask will trust any X-Forward-... header.
CondaStoreServer.template initialized
fastapi.templating.Jinja2Templates to use for html templates.
CondaStoreServer.template_vars extra variables to be passed into
jinja templates for page rendering.
CondaStoreServer.additional_routes additional routes for conda-store
to serve in form [(path, method, function), ...]. path is a
string, method is get, post, put, delete etc. and function
is a regular python fastapi function.
conda_store_server.worker.app.CondaStoreWorker#
CondaStoreWorker.log_level is the level for all server
logging. Default is INFO. Common options are DEBUG, INFO,
WARNING, and ERROR.
CondaStoreWorker.watch_paths is a list of paths for conda-store to
watch for changes to directories of environment.yaml files or a
single filename to watch.
CondaStoreWorker.concurrency by default is not set and defaults to
the number of threads on your given machine. If set will limit the
number of concurrent celery tasks to the integer.
conda_store_server.registry.ContainerRegistry#
ContainerRegistry.container_registries dictionary of registries_url
to upload built container images with callable function to configure
registry instance with credentials. Example configuration shown
bellow. Some registries are more complex to setup such as ECR, GCR,
etc. password is often the token generated from the AWS, GCP, Azure,
and Digital Ocean clients.
from python_docker.registry import Registry
import os
def _configure_docker_registry(registry_url: str):
return Registry(
"https://registry-1.docker.io",
username=os.environ.get('DOCKER_USERNAME'),
password=os.environ.get('DOCKER_PASSWORD'))
c.ContainerRegistry.container_registries = {
'https://registry-1.docker.io': _configure_docker_registry
}
ContainerRegistry.container_registry_image_name image name
to assign to docker image pushed for particular registry via a
callable function with arguments of registry and build.
ContainerRegistry.container_registry_image_tag image tag
to assign to docker image pushed for particular registry via a
callable function with arguments of registry and build.
Frequently Asked Questions#
conda-store fails to build Conda environment and worker is spontaneously killed (9 SIGKILL)#
The following error most likely indicates that you have not allocated
enough memory to conda-store-workers for solving and building the
given environment. Solve this by increasing the memory allocated to
the container.
Process 'ForkPoolWorker-31' pid:90 exited with 'signal 9 (SIGKILL)'
Task handler raised error: WorkerLostError('Worker exited prematurely: signal 9 (SIGKILL) Job: 348.')
Traceback (most recent call last):
File "/opt/conda/envs/conda-store-server/lib/python3.9/site-packages/billiard/pool.py", line 1265, in mark_as_worker_lost
raise WorkerLostError(
billiard.exceptions.WorkerLostError: Worker exited prematurely: signal 9 (SIGKILL) Job: 348.
Why are environment builds stuck in building state?#
Recently conda-store added a feature to cleanup builds which are stuck in the BUILDING state and are not currently running on the workers. This feature only works for certain brokers e.g. redis. Database celery brokers are not supported.
This issue occurs when the worker spontaineously dies. This can happen for several reasons:
worker is killed due to consuming too much memory (conda solver/builds can consume a lot of memory)
worker was killed for other reasons e.g. forced restart
bugs in conda-store