dallinger/experiment.py
"""The base experiment class."""
from __future__ import print_function, unicode_literals
import datetime
import inspect
import json
import logging
import os
import random
import sys
import time
import uuid
import warnings
from collections import Counter, OrderedDict
from contextlib import contextmanager
from functools import wraps
from importlib import import_module
from operator import itemgetter
from typing import List, Optional, Union
import requests
from cached_property import cached_property
from flask import Blueprint
from sqlalchemy import Table, and_, create_engine, func
from sqlalchemy.orm import scoped_session, sessionmaker, undefer
from sqlalchemy.orm.exc import MultipleResultsFound, NoResultFound
from dallinger import db, models, recruiters
from dallinger.config import LOCAL_CONFIG, get_config, initialize_experiment_package
from dallinger.data import (
Data,
export,
find_experiment_export,
ingest_zip,
is_registered,
)
from dallinger.data import load as data_load
from dallinger.db import (
Base,
db_url,
get_mapped_class,
get_polymorphic_mapping,
init_db,
)
from dallinger.heroku.tools import HerokuApp
from dallinger.information import Gene, Meme, State
from dallinger.models import Info, Network, Node, Participant, Transformation
from dallinger.networks import Empty
from dallinger.nodes import Agent, Environment, Source
from dallinger.transformations import Compression, Mutation, Replication, Response
from dallinger.utils import deferred_route_decorator, struct_to_html
logger = logging.getLogger(__file__)
def exp_class_working_dir(meth):
@wraps(meth)
def new_meth(self, *args, **kwargs):
try:
config = get_config()
orig_path = os.getcwd()
new_path = os.path.dirname(sys.modules[self.__class__.__module__].__file__)
os.chdir(new_path)
# Override configs
config.register_extra_parameters()
config.load_from_file(LOCAL_CONFIG)
return meth(self, *args, **kwargs)
finally:
config.clear()
os.chdir(orig_path)
return new_meth
class Experiment(object):
"""Define the structure of an experiment."""
app_id = None
exp_config = None
replay_path = "/"
#: Optional Redis channel to subscribe to on launch. Note that
#: if you set the channel, you will probably also want to override the
#: :func:`~dallinger.experiment.Experiment.send` method, since this
#: is where messages from Redis will be consumed. Setting a value here
#: will also result in the experiment being subscribed to the
#: ``dallinger_control`` channel for messages related to
#: socket/subscription updates. This is also the default ``channel_name``
#: for messages sent using the
#: :func:`~dallinger.experiment.Experiment.publish_to_subscribers` method.
channel = None
#: Constructor for Participant objects. Callable returning an instance of
#: :attr:`~dallinger.models.Participant` or a sub-class. Used by
#: :func:`~dallinger.experiment.Experiment.create_participant`.
participant_constructor = Participant
#: Flask Blueprint for experiment. Functions and methods on the class
#: should be registered as Flask routes using the
#: :func:`~dallinger.experiment.experiment_route` decorator. Route
#: functions can not be instance methods and should either be
#: plain functions or classmethods. You can also register route functions
#: at the module level using the standard `route` decorator on this
#: Blueprint.
experiment_routes = Blueprint(
"experiment_routes",
__name__,
template_folder="templates",
static_folder="static",
)
#: Sequence of dashboard route/function names that should be excluded from
#: rendering as tabs in the dashboard view.
hidden_dashboards = ()
def __init__(self, session=None):
"""Create the experiment class. Sets the default value of attributes."""
#: Boolean, determines whether the experiment logs output when
#: running. Default is True.
self.verbose = True
#: String, the name of the experiment. Default is "Experiment
#: title".
self.task = "Experiment title"
#: session, the experiment's connection to the database.
self.session = session
#: int, the number of practice networks (see
#: :attr:`~dallinger.models.Network.role`). Default is 0.
self.practice_repeats = 0
#: int, the number of non practice networks (see
#: :attr:`~dallinger.models.Network.role`). Default is 0.
self.experiment_repeats = 0
#: int, the number of participants
#: required to move from the waiting room to the experiment.
#: Default is 0 (no waiting room).
self.quorum = 0
#: int, the number of participants
#: requested when the experiment first starts. Default is 1.
self.initial_recruitment_size = 1
#: dictionary, the classes Dallinger can make in response
#: to front-end requests. Experiments can add new classes to this
#: dictionary.
self.known_classes = {
"Agent": Agent,
"Compression": Compression,
"Environment": Environment,
"Gene": Gene,
"Info": Info,
"Meme": Meme,
"Mutation": Mutation,
"Node": Node,
"Replication": Replication,
"Response": Response,
"Source": Source,
"State": State,
"Transformation": Transformation,
}
#: dictionary, the properties of this experiment that are exposed
#: to the public over an AJAX call
if not hasattr(self, "public_properties"):
# Guard against subclasses replacing this with a @property
self.public_properties = {}
if session:
self.configure()
try:
location = type(self).__module__
parent, experiment_module = location.rsplit(".", 1)
module = import_module(parent + ".jupyter")
except (ImportError, ValueError):
try:
from .jupyter import ExperimentWidget
self.widget = ExperimentWidget(self)
except ImportError:
self.widget = None
else:
self.widget = module.ExperimentWidget(self)
@staticmethod
def before_request():
return None
@staticmethod
def after_request(request, response):
return response
@classmethod
def get_status(cls):
"""
Return the status of the experiment as a dictionary.
"""
n_working_participants = (
db.session.query(func.count(Participant.id))
.filter_by(status="working")
.scalar()
)
return {"n_working_participants": n_working_participants}
@classmethod
def config_class(cls):
"""
Override this method in order to define a custom Configuration class
for dealing with config variables (see e.g. config.txt).
"""
from .config import Configuration
return Configuration
@classmethod
def extra_parameters(cls):
"""Override this classmethod to register new config variables. It is
called during config load. See
:ref:`Extra Configuration <extra-configuration>` for an example.
"""
pass
@classmethod
def config_defaults(cls):
"""Override this classmethod to register new default values for config variables."""
return {}
@property
def protected_routes(self):
"""Disable one or more standard Dallinger Flask routes by name.
When called, Flask routes which have been disabled will raise a
PermissionError and return a 500 response.
By default, this list is loaded from the `protected_routes` config parameter,
and is parsed as a JSON array. The values should be route rule names,
like "/" for the application root, or "/info/<int:node_id>/<int:info_id>"
for fetching JSON for a specific `Info`.
"""
return json.loads(get_config().get("protected_routes", "[]"))
def configure(self):
"""Load experiment configuration here"""
pass
@property
def background_tasks(self):
"""An experiment may define functions or methods to be started as
background tasks upon experiment launch.
"""
return []
def on_launch(self):
"""This function is called upon experiment launch. Unlike
the background tasks, this function is blocking: recruitment
won't start until the function has returned.
"""
pass
@cached_property
def recruiter(self):
"""Reference to a Recruiter, the Dallinger class that recruits
participants.
"""
return recruiters.from_config(get_config())
def calculate_qualifications(self, participant):
"""All the qualifications we want to assign to a worker.
This default implementation produces qualifications compatible with
Dallinger's standard recruiters, and the MTurkRecruiter in particular.
Workers will always be assigned one qualification specific to the
experiment run. If a "group_name" config value is set, this will be
parsed for additional qualifications to grant.
Return type is a list of dictionaries with "name", "description", and
optionally "score" (an integer), or an empty list.
"""
experiment_qualification_desc = "Experiment-specific qualification"
group_qualification_desc = "Experiment group qualification"
config = get_config()
group_names = [
n.strip() for n in config.get("group_name", "").split(",") if n.strip()
]
# Experiment-run specific:
quals = [
{
"name": config.get("id"),
"description": experiment_qualification_desc,
}
]
# From group_name:
quals.extend(
[
{"name": name, "description": group_qualification_desc}
for name in group_names
]
)
return quals
def is_overrecruited(self, waiting_count):
"""Returns True if the number of people waiting is in excess of the
total number expected, indicating that this and subsequent users should
skip the experiment. A quorum value of 0 means we don't limit
recruitment, and always return False.
"""
if not self.quorum:
return False
return waiting_count > self.quorum
def send(self, raw_message):
"""Async implementation of websocket message processing. Attempts to
extract a participant id or node id from the message, and send the
message to be processed asynchronously by
:func:`~dallinger.experiment.Experiment.receive_message` If it fails to
find a participant or node id in the message, then the message is
processed synchronously using
:func:`~dallinger.experiment.Experiment.receive_message`.
``raw_message`` is a string that includes a channel name prefix, for
example a JSON message for a ``shopping`` channel might look like:
``'shopping:{"type":"buy","color":"blue","quantity":"2"}'``
Control messages about channel subscription and websocket
connect/disconnect events use the ``dallinger_control`` channel name.
Experiments can override this method if they want to process all
messages synchronously in a single application instance by default. For
example if an experiment retains non-persisted state in an attribute of
the experiment class that it uses for message responses then it's best
to override this method instead of
:func:`~dallinger.experiment.Experiment.receive_message`, and explicitly
hand off state synchronization and other database writes to async worker
events.
:param raw_message: a formatted message string ``'$channel_name:$data'``
:type raw_message: str
"""
from dallinger.experiment_server.worker_events import (
_get_queue,
worker_function,
)
receive_time = datetime.datetime.now()
channel_name, message_string = raw_message.split(":", 1)
try:
message = json.loads(message_string)
except Exception:
# Not JSON we have no information about the participant/node and
# will run synchonously
self.receive_message(
message_string, channel_name=channel_name, receive_time=receive_time
)
return
participant_id = (
message.get("sender")
or message.get("participant_id")
or message.get("client", {}).get("participant_id")
)
node_id = message.get("node_id")
if not participant_id and not node_id:
self.receive_message(
message_string, channel_name=channel_name, receive_time=receive_time
)
return
q = _get_queue("high")
q.enqueue(
worker_function,
"WebSocketMessage",
None,
participant_id,
node_id=node_id,
receive_timestamp=receive_time.timestamp(),
details={
"message": message_string,
"channel_name": channel_name,
},
queue_name="high",
)
def receive_message(
self, message, channel_name=None, participant=None, node=None, receive_time=None
):
"""Stub implementation of a websocket message processor. Messages
are are queued to be processed asynchronously by
:func:`~dallinger.experiment.Experiment.send` and the worker runs this
method to process those messages. Sub-classes that wish to handle
incoming messages asynchronously should override this method. Generally
this method should always be overridden whenever the ``Experiment``
:attr:`~dallinger.experiment.Experiment.channel` attribute is set.
``message`` is a string, e.g. containing JSON formatted data.
Control messages about channel subscription and websocket
connect/disconnect events are sent over the ``"dallinger_control"``
channel.
This method is called synchronously when no participant or node
id can be determined from the message.
:param message: a websocket message
:type message: str
:param channel_name: The name of the channel the message was received on.
:type channel_name: str
:param participant: the experiment participant object responsible for the message
:type participant: :attr:`~dallinger.models.Participant` instance
:param node: the experiment node the message corresponds to
:type node: :attr:`~dallinger.models.Node` instance
:param receive_time: The time the message was received by the experiment
:type receive_time: datetime.datetime
"""
pass
def publish_to_subscribers(self, data, channel_name=None):
"""Publish data to the given channel_name. Data will be sent to all
channel subscribers, potentially including the experiment instance
itself. If no ``channel_name`` is specified, then the ``Experiment``
:attr:`~dallinger.experiment.Experiment.channel` value will be used
(and the data will automatically be consumed by
:func:`~dallinger.experiment.Experiment.send`). The ``data`` must be
a string, it typically contains JSON.
:param data: the message data to be send
:type data: str
:param channel_name: the name of the channel to publish the data to
:type channel_name: str
"""
if channel_name is None:
channel_name = self.channel
db.redis_conn.publish(channel_name, data)
def client_info(self):
"""Returns a JSON compatible dictionary with data about this client to
be included in control channel messages.
"""
return {
"class": self.__class__.__module__ + "." + self.__class__.__name__,
}
def setup(self):
"""Create the networks if they don't already exist."""
if not self.networks():
for _ in range(self.practice_repeats):
network = self.create_network()
network.role = "practice"
self.session.add(network)
for _ in range(self.experiment_repeats):
network = self.create_network()
network.role = "experiment"
self.session.add(network)
self.session.commit()
def create_network(self):
"""Return a new network."""
return Empty()
def networks(self, role="all", full="all"):
"""All the networks in the experiment."""
if full not in ["all", True, False]:
raise ValueError(
"full must be boolean or all, it cannot be {}".format(full)
)
if full == "all":
if role == "all":
return Network.query.all()
else:
return Network.query.filter_by(role=role).all()
else:
if role == "all":
return Network.query.filter_by(full=full).all()
else:
return Network.query.filter(
and_(Network.role == role, Network.full == full)
).all()
def get_network_for_participant(self, participant):
"""Find a network for a participant.
If no networks are available, None will be returned. By default
participants can participate only once in each network and participants
first complete networks with `role="practice"` before doing all other
networks in a random order.
"""
key = participant.id
networks_with_space = (
Network.query.filter_by(full=False).order_by(Network.id).all()
)
networks_participated_in = [
node.network_id
for node in Node.query.with_entities(Node.network_id)
.filter_by(participant_id=participant.id)
.all()
]
legal_networks = [
net for net in networks_with_space if net.id not in networks_participated_in
]
if not legal_networks:
self.log("No networks available, returning None", key)
return None
self.log(
"{} networks out of {} available".format(
len(legal_networks), (self.practice_repeats + self.experiment_repeats)
),
key,
)
legal_practice_networks = [
net for net in legal_networks if net.role == "practice"
]
if legal_practice_networks:
chosen_network = legal_practice_networks[0]
self.log(
"Practice networks available."
"Assigning participant to practice network {}.".format(
chosen_network.id
),
key,
)
else:
chosen_network = self.choose_network(legal_networks, participant)
self.log(
"No practice networks available."
"Assigning participant to experiment network {}".format(
chosen_network.id
),
key,
)
return chosen_network
def choose_network(self, networks, participant):
return random.choice(networks)
def create_node(self, participant, network):
"""Create a node for a participant."""
return Node(network=network, participant=participant)
def add_node_to_network(self, node, network):
"""Add a node to a network.
This passes `node` to :func:`~dallinger.models.Network.add_node()`.
"""
network.add_node(node)
def normalize_entry_information(self, entry_information):
"""Accepts a dictionary with information about a recruited user. Returns
a dictionary containing data the needed to create or load a Dallinger
Participant. The returned data should include valid ``assignment_id``,
``worker_id``, and ``hit_id`` values. It may also include an
``entry_information`` key which should contain a transformed
``entry_information`` dict which will be stored for newly created
participants.
By default, the extraction of these values is delegated to the
recruiter's `normalize_entry_information` method.
Returning a dictionary without valid ``hit_id``, ``assignment_id``, or
``worker_id`` will generally result in an exception.
"""
entry_data = self.recruiter.normalize_entry_information(entry_information)
# We need an assignment_id in order to create a participant
return entry_data
def create_participant(
self,
worker_id,
hit_id,
assignment_id,
mode,
recruiter_name=None,
fingerprint_hash=None,
entry_information=None,
):
"""Creates and returns a new participant object. Uses
:attr:`~dallinger.experiment.Experiment.participant_constructor` as the
constructor.
:param worker_id: the recruiter Worker Id
:type worker_id: str
:param hit_id: the recruiter HIT Id
:type hit_id: str
:param assignment_id: the recruiter Assignment Id
:type assignment_id: str
:param mode: the application mode
:type mode: str
:param recruiter_name: the recruiter name
:type recruiter_name: str
:param fingerprint_hash: the user's fingerprint
:type fingerprint_hash: str
:param entry_information: a JSON serializable data structure containing
additional participant entry information
:returns: A :attr:`~dallinger.models.Participant` instance
"""
if not recruiter_name:
recruiter = self.recruiter
if recruiter:
recruiter_name = recruiter.nickname
participant = self.participant_constructor(
recruiter_id=recruiter_name,
worker_id=worker_id,
assignment_id=assignment_id,
hit_id=hit_id,
mode=mode,
fingerprint_hash=fingerprint_hash,
entry_information=entry_information,
)
self.session.add(participant)
return participant
def load_participant(self, assignment_id):
"""Returns a participant object looked up by assignment_id.
Intended to allow a user to resume a session in a running experiment.
:param assignment_id: the recruiter Assignment Id
:type assignment_id: str
:returns: A ``Participant`` instance or ``None`` if there is not a
single matching participant.
"""
try:
return Participant.query.filter_by(assignment_id=assignment_id).one()
except (NoResultFound, MultipleResultsFound):
return None
def data_check(self, participant):
"""Check that the data are acceptable.
Return a boolean value indicating whether the `participant`'s data is
acceptable. This is meant to check for missing or invalid data. This
check will be run once the `participant` completes the experiment. By
default performs no checks and returns True. See also,
:func:`~dallinger.experiments.Experiment.attention_check`.
"""
return True
def bonus(self, participant):
"""The bonus to be awarded to the given participant.
Return the value of the bonus to be paid to `participant`. By default
returns 0.
"""
return 0
def bonus_reason(self):
"""The reason offered to the participant for giving the bonus.
Return a string that will be included in an email sent to the
`participant` receiving a bonus. By default it is "Thank you for
participating! Here is your bonus."
"""
return "Thank for participating! Here is your bonus."
def exit_info_for(self, participant):
"""An experiment can return a dictionary of infomation that will
be shown to the participant at the very last point in their
lifecycle, if the HIT is not submitted to an external recruitment
service for submission.
For complete control over the exit page, a customized version of
the ``exit_recruiter.html`` template can be included in the experient
directory, and this will override the default provided by Dallinger.
:param participant: the ``Participant`` instance for which to calculate
an exit value
:returns: ``dict`` which may be rendered to the worker as an HTML table
when they submit their assigment.
"""
return {
"Assignment ID": participant.assignment_id,
"HIT ID": participant.hit_id,
"Base Pay": participant.base_pay,
"Bonus": participant.bonus,
}
def attention_check(self, participant):
"""Check if participant performed adequately.
Return a boolean value indicating whether the `participant`'s data is
acceptable. This is mean to check the participant's data to determine
that they paid attention. This check will run once the *participant*
completes the experiment. By default performs no checks and returns
True. See also :func:`~dallinger.experiments.Experiment.data_check`.
"""
return True
def on_assignment_submitted_to_recruiter(self, participant, event):
"""Working title. Called when assignment has been submitted
to the recruitment platform (may be Dallinger itself) by the
participant.
:param participant (Participant): the ``Participant`` who has
submitted a HIT via their recruiter
:param event: (dict): Info about the triggering event
"""
eligible_statuses = ("working", "overrecruited", "returned", "abandoned")
if participant.status not in eligible_statuses:
return
config = get_config()
min_real_bonus = 0.01
participant.status = "submitted"
participant.end_time = event["timestamp"]
participant.base_pay = config.get("base_payment")
participant.recruiter.approve_hit(participant.assignment_id)
# Data Check
if not self.data_check(participant=participant):
participant.status = "bad_data"
self.data_check_failed(participant=participant)
# NB: if MultiRecruiter is in use, this may not be the same recruiter as
# provided the participant we're replacing
self.recruiter.recruit(n=1)
# NOTE EARLY RETURN!!
return
# If they pass the data check, we might pay a bonus
bonus = self.bonus(participant=participant)
participant.bonus = bonus
if bonus >= min_real_bonus:
self.log("Bonus = {}: paying bonus".format(bonus))
participant.recruiter.reward_bonus(
participant,
bonus,
self.bonus_reason(),
)
else:
self.log("Bonus = {}: NOT paying bonus".format(bonus))
# Attention Check
if self.attention_check(participant=participant):
self.log("Attention checks passed.")
participant.status = "approved"
self.submission_successful(participant=participant)
self.recruit()
else:
self.log("Attention checks failed.")
participant.status = "did_not_attend"
self.attention_check_failed(participant=participant)
# NB: if MultiRecruiter is in use, this may not be the same recruiter
# that provided the participant we're replacing:
self.recruiter.recruit(n=1)
def participant_task_completed(self, participant):
"""Called when an experiment task is finished and submitted, and prior
to data and attendance checks.
Assigns the qualifications to the Participant, via their recruiter.
These will include one Qualification for the experiment
ID, and others for the configured group_name, if it's been set.
Overrecruited participants don't receive qualifications, since they
haven't actually completed the experiment. This allows them to remain
eligible for future runs.
:param participant: the ``Participant`` instance
"""
config = get_config()
if not bool(config.get("assign_qualifications")):
logger.info("Qualification assignment is globally disabled; ignoring.")
return
if participant.status == "overrecruited":
return
quals = self.calculate_qualifications(participant)
participant.recruiter.assign_experiment_qualifications(
worker_id=participant.worker_id, qualifications=quals
)
def submission_successful(self, participant):
"""Run when a participant's experiment submission passes data
and attendence checks.
:param participant: the ``Participant`` instance
"""
pass
def recruit(self):
"""Recruit participants to the experiment as needed.
This method runs whenever a participant successfully completes the
experiment (participants who fail to finish successfully are
automatically replaced). By default it recruits 1 participant at a time
until all networks are full.
"""
if not self.networks(full=False):
self.log("All networks full: closing recruitment", "-----")
self.recruiter.close_recruitment()
def log(self, text, key="?????", force=False):
"""Print a string to the logs."""
if force or self.verbose:
print(">>>> {} {}".format(key, text))
sys.stdout.flush()
def log_summary(self):
"""Log a summary of all the participants' status codes."""
participants = Participant.query.with_entities(Participant.status).all()
counts = Counter([p.status for p in participants])
sorted_counts = sorted(counts.items(), key=itemgetter(0))
self.log("Status summary: {}".format(str(sorted_counts)))
return sorted_counts
def save(self, *objects):
"""Add all the objects to the session and commit them.
This only needs to be done for networks and participants.
"""
if len(objects) > 0:
self.session.add_all(objects)
self.session.commit()
def node_post_request(self, participant, node):
"""Run when a request to make a node is complete."""
pass
def node_get_request(self, node=None, nodes=None):
"""Run when a request to get nodes is complete."""
pass
def vector_post_request(self, node, vectors):
"""Run when a request to connect is complete."""
pass
def vector_get_request(self, node, vectors):
"""Run when a request to get vectors is complete."""
pass
def info_post_request(self, node, info):
"""Run when a request to create an info is complete."""
pass
def info_get_request(self, node, infos):
"""Run when a request to get infos is complete."""
pass
def transmission_post_request(self, node, transmissions):
"""Run when a request to transmit is complete."""
pass
def transmission_get_request(self, node, transmissions):
"""Run when a request to get transmissions is complete."""
pass
def transformation_post_request(self, node, transformation):
"""Run when a request to transform an info is complete."""
pass
def transformation_get_request(self, node, transformations):
"""Run when a request to get transformations is complete."""
pass
def fail_participant(self, participant):
"""Fail all the nodes of a participant."""
participant_nodes = Node.query.filter_by(
participant_id=participant.id, failed=False
).all()
for node in participant_nodes:
node.fail()
def data_check_failed(self, participant):
"""What to do if a participant fails the data check.
Runs when `participant` has failed
:func:`~dallinger.experiments.Experiment.data_check`. By default calls
:func:`~dallinger.experiments.Experiment.fail_participant`.
"""
self.fail_participant(participant)
def attention_check_failed(self, participant):
"""What to do if a participant fails the attention check.
Runs when `participant` has failed the
:func:`~dallinger.experiments.Experiment.attention_check`. By default calls
:func:`~dallinger.experiments.Experiment.fail_participant`.
"""
self.fail_participant(participant)
def assignment_abandoned(self, participant):
"""What to do if a participant abandons the hit.
This runs when a notification from AWS is received indicating that
`participant` has run out of time. Calls
:func:`~dallinger.experiments.Experiment.fail_participant`.
"""
self.fail_participant(participant)
def assignment_returned(self, participant):
"""What to do if a participant returns the hit.
This runs when a notification from AWS is received indicating that
`participant` has returned the experiment assignment. Calls
:func:`~dallinger.experiments.Experiment.fail_participant`.
"""
self.fail_participant(participant)
def assignment_reassigned(self, participant):
"""What to do if the assignment assigned to a participant is
reassigned to another participant while the first participant
is still working.
This runs when a participant is created with the same assignment_id
as another participant if the earlier participant still has the status
"working". Calls :func:`~dallinger.experiments.Experiment.fail_participant`.
"""
self.fail_participant(participant)
@exp_class_working_dir
def run(self, exp_config=None, app_id=None, bot=False, **kwargs):
"""Deploy and run an experiment.
The exp_config object is either a dictionary or a
``localconfig.LocalConfig`` object with parameters
specific to the experiment run grouped by section.
"""
import dallinger as dlgr
app_id = self.make_uuid(app_id)
if bot:
kwargs["recruiter"] = "bots"
self.app_id = app_id
self.exp_config = exp_config or kwargs
self.update_status("Starting")
try:
if self.exp_config.get("mode") == "debug":
dlgr.command_line.debug.callback(
verbose=True, bot=bot, proxy=None, exp_config=self.exp_config
)
else:
dlgr.deployment.deploy_sandbox_shared_setup(
dlgr.command_line.log,
app=app_id,
verbose=self.verbose,
exp_config=self.exp_config,
)
except Exception:
self.update_status("Errored")
raise
else:
self.update_status("Running")
self._await_completion()
self.update_status("Retrieving data")
data = self.retrieve_data()
self.update_status("Completed")
return data
def collect(self, app_id, exp_config=None, bot=False, **kwargs):
"""Collect data for the provided experiment id.
The ``app_id`` parameter must be a valid UUID.
If an existing data file is found for the UUID it will
be returned, otherwise - if the UUID is not already registered -
the experiment will be run and data collected.
See :meth:`~Experiment.run` method for other parameters.
"""
try:
results = data_load(app_id)
self.log(
"Data found for experiment {}, retrieving.".format(app_id),
key="Retrieve:",
)
return results
except IOError:
self.log(
"Could not fetch data for id: {}, checking registry".format(app_id),
key="Retrieve:",
)
exp_config = exp_config or {}
if is_registered(app_id):
raise RuntimeError(
"The id {} is registered, ".format(app_id)
+ "but you do not have permission to access to the data"
)
elif kwargs.get("mode") == "debug" or exp_config.get("mode") == "debug":
raise RuntimeError("No remote or local data found for id {}".format(app_id))
try:
assert isinstance(uuid.UUID(app_id, version=4), uuid.UUID)
except (ValueError, AssertionError):
raise ValueError("Invalid UUID supplied {}".format(app_id))
self.log(
"{} appears to be a new experiment id, running experiment.".format(app_id),
key="Retrieve:",
)
return self.run(exp_config, app_id, bot, **kwargs)
@classmethod
def make_uuid(cls, app_id=None):
"""Generates a new UUID.
This is a class method and can be called as `Experiment.make_uuid()`.
Takes an optional `app_id` which is converted to a string and, if it
is a valid UUID, returned.
"""
try:
if app_id and isinstance(uuid.UUID(str(app_id), version=4), uuid.UUID):
return str(app_id)
except (ValueError, AssertionError):
pass
return str(uuid.UUID(int=random.getrandbits(128)))
def experiment_completed(self):
"""Checks the current state of the experiment to see whether it has
completed. This makes use of the experiment server `/summary` route,
which in turn uses :meth:`~Experiment.is_complete`.
"""
heroku_app = HerokuApp(self.app_id)
status_url = "{}/summary".format(heroku_app.url)
data = {}
try:
resp = requests.get(status_url)
data = resp.json()
except (ValueError, requests.exceptions.RequestException):
logger.exception("Error fetching experiment status.")
logger.debug("Current application state: {}".format(data))
return data.get("completed", False)
def _await_completion(self):
# Debug runs synchronously, but in live mode we need to loop and check
# experiment status
if self.exp_config.get("mode") != "debug":
self.log("Waiting for experiment to complete.", "")
while not self.experiment_completed():
time.sleep(30)
return True
def retrieve_data(self):
"""Retrieves and saves data from a running experiment"""
local = False
if self.exp_config.get("mode") == "debug":
local = True
filename = export(self.app_id, local=local)
logger.debug("Data exported to %s" % filename)
return Data(filename)
def end_experiment(self):
"""Terminates a running experiment"""
if self.exp_config.get("mode") != "debug":
HerokuApp(self.app_id).destroy()
return True
def events_for_replay(self, session=None, target=None):
"""Returns an ordered list of "events" for replaying.
Experiments may override this method to provide custom
replay logic. The "events" returned by this method will be passed
to :meth:`~Experiment.replay_event`. The default implementation
simply returns all :class:`~dallinger.models.Info` objects in the
order they were created.
"""
if session is None:
session = self.session
return session.query(Info).order_by(Info.creation_time)
def replay_event(self, event):
"""Stub method to replay an event returned by
:meth:`~Experiment.events_for_replay`.
Experiments must override this method to provide replay support.
"""
pass
def replay_start(self):
"""Stub method for starting an experiment replay.
Experiments must override this method to provide replay support.
"""
pass
def replay_finish(self):
"""Stub method for ending an experiment replay.
Experiments must override this method to provide replay support.
"""
pass
def replay_started(self):
"""Returns `True` if an experiment replay has started."""
return True
def is_complete(self):
"""Method for custom determination of experiment completion.
Experiments should override this to provide custom experiment
completion logic. Returns `None` to use the experiment server
default logic, otherwise should return `True` or `False`.
"""
return None
def monitoring_panels(self, **kw):
"""Provides monitoring dashboard sidebar panels.
:param \\**kw: arguments passed in from the request
:returns: An ``OrderedDict()`` mapping panel titles to HTML strings
to render in the dashboard sidebar.
""" # noqa
stats = self.monitoring_statistics(**kw)
panels = OrderedDict()
for tab in stats:
panels[tab] = struct_to_html(stats[tab])
return panels
def monitoring_statistics(self, **kw):
"""The default data used for the monitoring panels
:param \\**kw: arguments passed in from the request
:returns: An ``OrderedDict()`` mapping panel titles to data structures
describing the experiment state.
""" # noqa
participants = Participant.query
nodes = Node.query
infos = Info.query
stats = OrderedDict()
stats["Participants"] = OrderedDict(
(
("working", participants.filter_by(status="working").count()),
("abandoned", participants.filter_by(status="abandoned").count()),
("returned", participants.filter_by(status="returned").count()),
("approved", participants.filter_by(status="approved").count()),
)
)
# Count up our networks by role
network_roles = self.session.query(Network.role, func.count(Network.role))
network_counts = network_roles.group_by(Network.role).all()
failed_networks = network_roles.filter(Network.failed == True) # noqa
failed_counts = dict(failed_networks.group_by(Network.role).all())
network_stats = {}
for role, count in network_counts:
network_stats[role] = OrderedDict(
(
("count", count),
("failed", failed_counts.get(role, 0)),
)
)
stats["Networks"] = network_stats
stats["Nodes"] = OrderedDict(
(
("count", nodes.count()),
("failed", nodes.filter_by(failed=True).count()),
)
)
stats["Infos"] = OrderedDict(
(
("count", infos.count()),
("failed", infos.filter_by(failed=True).count()),
)
)
if kw.get("transformations"):
transformations = Transformation.query
stats["transformations"] = OrderedDict(
(
("count", transformations.count()),
("failed", transformations.filter_by(failed=True).count()),
)
)
return stats
def network_structure(
self,
network_roles=None,
network_ids=None,
collapsed=False,
transformations=False,
):
networks = self.summarize_table("network", network_roles, network_ids)
nodes = self.summarize_table(
"node",
network_roles,
network_ids,
cls_filter=(lambda cls: issubclass(cls, Source)) if collapsed else None,
)
if collapsed:
vectors = []
infos = []
participants = []
trans = []
else:
vectors = self.summarize_table("vector", network_roles, network_ids)
infos = self.summarize_table("info", network_roles, network_ids)
participants = self.summarize_table("participant")
if transformations:
trans = self.summarize_table(
"transformation", network_roles, network_ids
)
else:
trans = []
return {
"networks": networks,
"nodes": nodes,
"vectors": vectors,
"infos": infos,
"participants": participants,
"trans": trans,
}
def summarize_table(
self,
table: Union[Table, str],
network_roles: Optional[List] = None,
network_ids: Optional[List] = None,
cls_filter: Optional[callable] = None,
):
"""
Summarizes a given database table.
:param table: Table to be summarized
:param network_roles: Optionally restrict output to objects from networks with these roles
:param network_ids: Optionally restrict output to objects from networks with these IDs
:param cls_filter: Optional lambda function that returns ``False`` for classes that should be excluded
Returns a list of JSON-style dictionaries produced by calling ``.__json__()`` on every object
retrieved from the table.
"""
objects = self.pull_table(
table=table,
polymorphic_identity=None,
network_roles=network_roles,
network_ids=network_ids,
cls_filter=cls_filter,
)
return [obj.__json__() for obj in objects]
def pull_table(
self,
table: Union[Table, str],
polymorphic_identity: Optional[str] = None,
network_roles: Optional[List] = None,
network_ids: Optional[List] = None,
cls_filter: Optional[callable] = None,
):
"""
Downloads every object in the specified table.
For efficiency, the SQL queries are batched by the values of the polymorphic identity column ``type``
if it is present.
:param table: Table to be summarized
:param polymorphic_identity: Optionally restrict output to a given polymorphic identity (i.e. ``type`` value)
:param network_roles: Optionally restrict output to objects from networks with these roles
:param network_ids: Optionally restrict output to objects from networks with these IDs
:param cls_filter: Optional lambda function that returns ``False`` for classes that should be excluded
Returns a list of database-mapped objects.
"""
if isinstance(table, str):
table = Base.metadata.tables[table]
if polymorphic_identity is None and "type" in table.columns:
observed_types = [
r.type for r in db.session.query(table.columns.type).distinct().all()
]
obj_by_type = [
self.pull_table(
table,
polymorphic_identity=_type,
network_roles=network_roles,
network_ids=network_ids,
cls_filter=cls_filter,
)
for _type in observed_types
]
return [obj for sublist in obj_by_type for obj in sublist]
if polymorphic_identity is None:
cls = get_mapped_class(table)
else:
assert "type" in table.columns
cls = get_polymorphic_mapping(table)[polymorphic_identity]
if cls_filter is not None and not cls_filter(cls):
return
query = cls.query
if polymorphic_identity is not None:
query = query.filter(cls.type == polymorphic_identity)
if network_roles is not None:
query = query.filter(Network.role.in_(network_roles))
if network_ids is not None:
query = query.filter(Network.id.in_(network_ids))
if network_roles is not None or network_ids is not None:
if "network_id" in table.columns:
query = query.join(Network, cls.network_id == Network.id)
primary_keys = [c.name for c in table.primary_key.columns]
return query.order_by(*primary_keys).options(undefer("*")).all()
def node_visualization_options(self):
"""Provides custom vis.js configuration options for the
Network Monitoring Dashboard.
:returns: A dict with `vis.js option values <https://visjs.github.io/vis-network/docs/network/#options>`__
"""
return {}
def node_visualization_html(self, object_type, obj_id):
"""Returns a string with custom HTML visualization for a given object
referenced by the object base type and id.
:param object_type: The base object class name, e.g. ``Network``, ``Node``, ``Info``, ``Participant``, etc.
:type object_type: str
:param id: The ``id`` of the object
:type id: int
:returns: A valid HTML string to be inserted into the monitoring dashboard
"""
model = getattr(models, object_type, None)
if model is not None:
obj = self.session.query(model).get(int(obj_id))
if getattr(obj, "visualization_html", None):
return obj.visualization_html
return ""
def table_data(
self, table: str = "participant", polymorphic_identity: Optional[str] = None
):
"""Generates DataTablesJS data and configuration for the experiment. The data
is compiled from the models' ``__json__`` methods, and can be customized by either
overriding this method or using the ``json_data`` method on the model to return
additional serializable data.
:param table: table to query
:param polymorphic_identity: optional polymorphic identity (corresponds to the ``type`` column)
:returns: Returns a ``dict`` with DataTablesJS data and configuration, filters using
arbitrary keyword arguments. Should contain ``data`` and ``columns`` keys
at least, with ``columns`` containing data for all fields on all returned
objects.
""" # noqa
rows = []
found_columns = set()
columns = []
table = Base.metadata.tables[table]
if polymorphic_identity == "None":
polymorphic_identity = None
objects = self.pull_table(table, polymorphic_identity=polymorphic_identity)
for obj in objects:
data = obj.__json__()
# Add participant worker_id to data, we normally leave it out of
# JSON renderings
if table.name == "participant":
data["worker_id"] = obj.worker_id
rows.append(data)
for key in data:
if key not in found_columns:
columns.append({"name": key, "data": key})
found_columns.add(key)
# Make sure every row has an entry for every column
for col in found_columns:
for row in rows:
if col not in row:
row[col] = None
return {
"data": rows,
"columns": columns,
}
def dashboard_database_actions(self):
"""Returns a sequence of custom actions for the database dashboard. Each action
must have a ``title`` and a ``name`` corresponding to a method on the
experiment class.
The named methods should take a single ``data`` argument
which will be a list of dicts representing the datatables rendering of
a Dallinger model object. The named methods should return a ``dict``
containing a ``"message"`` which will be displayed in the dashboard.
Returns a single action referencing the
:func:`~dallinger.experiment.Experiment.dashboard_fail`
method by default.
"""
return [{"name": "dashboard_fail", "title": "Fail Selected"}]
def dashboard_fail(self, data):
"""Marks matching non-failed items as failed. Items are looked up by
``id`` and ``object_type`` (e.g. ``"Participant"``).
:param data: A list of dicts representing model items to be marked as failed.
Each must have an ``id`` and an ``object_type``
:type object_type: list
:returns: Returns a ``dict`` with a ``"message"`` string indicating how
many items were successfully marked as failed.
"""
counts = {}
for entry in data:
obj_id = entry.get("id")
object_type = entry.get("object_type")
model = getattr(models, object_type, None)
if model is not None:
obj = self.session.query(model).get(int(obj_id))
if obj is not None and not obj.failed:
obj.fail()
counts[object_type] = counts.get(object_type, 0) + 1
if not counts:
return {"message": "No nodes found to fail"}
return {
"message": "Failed {}".format(
", ".join("{} {}s".format(c, t) for t, c in sorted(counts.items()))
)
}
@property
def usable_replay_range(self):
"""The range of times that represent the active part of the experiment"""
return self._replay_range
@contextmanager
def restore_state_from_replay(
self, app_id, session, zip_path=None, **configuration_options
):
# We need to fake dallinger_experiment to point at the current experiment
module = sys.modules[type(self).__module__]
if sys.modules.get("dallinger_experiment", module) != module:
logger.warning("dallinger_experiment is already set, updating")
sys.modules["dallinger_experiment"] = module
# Load the configuration system and globals
config = get_config()
# Manually load extra parameters and ignore errors
try:
from dallinger_experiment.experiment import extra_parameters
try:
extra_parameters()
config._module_params_loaded = True
except KeyError:
pass
except ImportError:
pass
config.load()
self.app_id = self.original_app_id = app_id
self.session = session
self.exp_config = config
# The replay index is initialised to 1970 as that is guaranteed
# to be before any experiment Info objects
self._replay_time_index = datetime.datetime(1970, 1, 1, 1, 1, 1)
# Create a second database session so we can load the full history
# of the experiment to be replayed and selectively import events
# into the main database
specific_db_url = db_url + "-import-" + app_id
import_engine = create_engine(specific_db_url)
try:
# Clear the temporary storage and import it
init_db(drop_all=True, bind=import_engine)
except Exception:
create_db_engine = create_engine(db_url)
conn = create_db_engine.connect()
conn.execute("COMMIT;")
conn.execute(
'CREATE DATABASE "{}"'.format(specific_db_url.rsplit("/", 1)[1])
)
conn.close()
import_engine = create_engine(specific_db_url)
init_db(drop_all=True, bind=import_engine)
self.import_session = scoped_session(
sessionmaker(autocommit=False, autoflush=True, bind=import_engine)
)
# Find the real data for this experiment
if zip_path is None:
zip_path = find_experiment_export(app_id)
if zip_path is None:
msg = 'Dataset export for app id "{}" could not be found.'
raise IOError(msg.format(app_id))
print("Ingesting dataset from {}...".format(os.path.basename(zip_path)))
ingest_zip(zip_path, engine=import_engine)
self._replay_range = tuple(
self.import_session.query(
func.min(Info.creation_time), func.max(Info.creation_time)
)
)[0]
# We apply the configuration options we were given and yield
# the scrubber function into the context manager, so within the
# with experiment.restore_state_from_replay(...): block the configuration
# options are correctly set
with config.override(configuration_options, strict=True):
self.replay_start()
yield Scrubber(self, session=self.import_session)
self.replay_finish()
# Clear up global state
self.import_session.rollback()
self.import_session.close()
session.rollback()
session.close()
config._reset(register_defaults=True)
del sys.modules["dallinger_experiment"]
def revert_to_time(self, session, target):
# We do not support going back in time
raise NotImplementedError
def _ipython_display_(self):
"""Display Jupyter Notebook widget"""
from IPython.display import display
display(self.widget)
def update_status(self, status):
if self.widget is not None:
self.widget.status = status
def jupyter_replay(self, *args, **kwargs):
from IPython.display import display
from ipywidgets import widgets
try:
sys.modules["dallinger_experiment"]._jupyter_cleanup()
except (KeyError, AttributeError):
pass
replay = self.restore_state_from_replay(*args, **kwargs)
scrubber = replay.__enter__()
scrubber.build_widget()
replay_widget = widgets.VBox([self.widget, scrubber.widget])
# Scrub to start of experiment and re-render the main widget
scrubber(self.usable_replay_range[0])
self.widget.render()
display(replay_widget)
# Defer the cleanup until this function is re-called by
# keeping a copy of the function on the experiment module
# This allows us to effectively detect the cell being
# re-run as there doesn't seem to be a cleanup hook for widgets
# displayed as part of a cell that is being re-rendered
def _jupyter_cleanup():
replay.__exit__(None, None, None)
sys.modules["dallinger_experiment"]._jupyter_cleanup = _jupyter_cleanup
class Scrubber(object):
def __init__(self, experiment, session):
self.experiment = experiment
self.session = session
self.realtime = False
def __call__(self, time):
"""Scrub to a point in the experiment replay, given by time
which is a datetime object."""
if self.experiment._replay_time_index > time:
self.experiment.revert_to_time(session=self.session, target=time)
events = self.experiment.events_for_replay(
session=self.session, target=time
).all()
for event in events:
if event.creation_time <= self.experiment._replay_time_index:
# Skip events we've already handled
continue
if event.creation_time > time:
# Stop once we get future events
break
self.experiment.replay_event(event)
self.experiment._replay_time_index = event.creation_time
# Override app_id to allow exports to be created that don't
# overwrite the original dataset
self.experiment.app_id = "{}_{}".format(
self.experiment.original_app_id, time.isoformat()
)
def in_realtime(self, callback=None):
exp_start, exp_end = self.experiment.usable_replay_range
replay_offset = time.time()
current = self.experiment._replay_time_index
if current < exp_start:
current = exp_start
self.realtime = True
# Disable the scrubbing slider
self.widget.children[0].disabled = True
try:
while current < exp_end:
now = time.time()
seconds = now - replay_offset
current = current + datetime.timedelta(seconds=seconds)
self(current)
if callable(callback):
try:
callback()
except StopIteration:
return
replay_offset = now
finally:
self.realtime = False
self.widget.children[0].disabled = False
def build_widget(self):
from ipywidgets import widgets
start, end = self.experiment.usable_replay_range
options = []
current = start
while current <= end:
# Never display microseconds
options.append((current.replace(microsecond=0).time().isoformat(), current))
current += datetime.timedelta(seconds=1)
# But we need to keep microseconds in the first value, so we don't go before
# the experiment start when scrubbing backwards
current = current.replace(microsecond=0)
scrubber = widgets.SelectionSlider(
description="Current time",
options=options,
disabled=False,
continuous_update=False,
)
def advance(change):
if self.realtime:
# We're being driven in realtime, the advancement
# here is just to keep the UI in sync
return
old_status = self.experiment.widget.status
self.experiment.widget.status = "Updating"
self.experiment.widget.render()
self(change["new"])
self.experiment.widget.status = old_status
self.experiment.widget.render()
scrubber.observe(advance, "value")
def realtime_callback():
self.experiment.widget.render()
try:
scrubber.value = self.experiment._replay_time_index.replace(
microsecond=0
)
except Exception:
# The scrubber is an approximation of the current time, we shouldn't
# bail out if it can't be updated (for example at experiment bounds)
pass
if not self.realtime:
raise StopIteration()
play_button = widgets.ToggleButton(
value=False,
description="",
disabled=False,
tooltip="Play back in realtime",
icon="play",
)
def playback(change):
import threading
if change["new"]:
thread = threading.Thread(
target=self.in_realtime, kwargs={"callback": realtime_callback}
)
thread.start()
else:
self.realtime = False
play_button.observe(playback, "value")
self.widget = widgets.HBox(children=[scrubber, play_button])
return self.widget
def _ipython_display_(self):
"""Display Jupyter Notebook widget"""
from IPython.display import display
self.build_widget()
display(self.widget())
def is_experiment_class(cls):
return (
inspect.isclass(cls) and issubclass(cls, Experiment) and cls is not Experiment
)
def load():
"""Load the active experiment."""
first_err = second_err = None
initialize_experiment_package(os.getcwd())
try:
try:
from dallinger_experiment import experiment
except ImportError as e:
first_err = e
try:
from dallinger_experiment import dallinger_experiment as experiment
except ImportError as e:
second_err = e
import dallinger_experiment as experiment
classes = inspect.getmembers(experiment, is_experiment_class)
preferred_class = os.environ.get("EXPERIMENT_CLASS_NAME", None)
if preferred_class is not None:
try:
return dict(classes)[preferred_class]
except KeyError:
raise ImportError(
"No experiment named {} was found".format(preferred_class)
)
if len(classes) > 1:
for name, c in classes:
if "Experiment" in c.__bases__[0].__name__:
warnings.warn(
UserWarning(
"More than one potential experiment class found but no EXPERIMENT_CLASS_NAME environment variable. Picking {} from {}.".format(
name, [n for (n, cls) in classes]
)
),
stacklevel=3,
)
return c
raise ImportError(
"No direct experiment subclass found in {}".format(
[n for (n, cls) in classes]
)
)
elif len(classes) == 0:
logger.error("Error retrieving experiment class")
if not module_is_initialized(experiment):
logger.error(
"The experiment module is only partly initialized. Maybe you have a circular import?"
)
raise (
first_err
or second_err
or ImportError("No classes found in {}".format(experiment))
)
else:
return classes[0][1]
except ImportError:
logger.error("Could not import experiment.")
raise
def module_is_initialized(module):
"""
Checks whether a given module has been initialized by catching the AttributeError that happens when accessing
an unknown attribute within that module. This is a bit of a hack, but it seems to be the easiest
way of checking the modules initialization status.
"""
try:
module.abcdefghijklmnop123456789
except AttributeError as err:
if "partially initialized module" in str(err):
return False
return True
EXPERIMENT_TASK_REGISTRATIONS = []
def scheduled_task(trigger, **kwargs):
"""Creates a decorator to register experiment functions or classmethods as
tasks for the clock process. Accepts all
arguments for `apscheduler.schedulers.base.BaseSchedule.scheduled_job`
The task registration is deferred until clock server setup to allow tasks to be
overridden by subclasses.
:param trigger: an ``apscheduler`` trigger type. One of "interval", "cron",
or "date"
:param \\**kwargs: other arguments for `apscheduler.schedulers.base.BaseSchedule.scheduled_job`
generally used for trigger arguments to determine
the run interval.
:returns: A decorator to register methods from a class as scheduled tasks.
""" # noqa
registered_tasks = EXPERIMENT_TASK_REGISTRATIONS
scheduler_args = {
"trigger": trigger,
"kwargs": tuple(kwargs.items()),
}
return deferred_route_decorator(scheduler_args, registered_tasks)
EXPERIMENT_ROUTE_REGISTRATIONS = []
def experiment_route(rule, **kwargs):
"""Creates a decorator to register experiment functions or classmethods as
routes on the ``Experiment.experiment_routes`` blueprint. Accepts all
standard flask ``route`` arguments. The registration is deferred until
experiment server setup to allow routes to be overridden.
:returns: A decorator to register methods from a class as experiment routes.
"""
registered_routes = EXPERIMENT_ROUTE_REGISTRATIONS
route = {
"rule": rule,
"kwargs": tuple(kwargs.items()),
}
return deferred_route_decorator(route, registered_routes)