piker/piker/data/_sampling.py

872 lines
29 KiB
Python

# piker: trading gear for hackers
# Copyright (C) 2018-present Tyler Goodlet (in stewardship of pikers)
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
"""
Sampling and broadcast machinery for (soft) real-time delivery of
financial data flows.
"""
from __future__ import annotations
from collections import (
Counter,
defaultdict,
)
from contextlib import asynccontextmanager as acm
import time
from typing import (
AsyncIterator,
TYPE_CHECKING,
)
import tractor
from tractor.trionics import (
maybe_open_nursery,
)
import trio
from trio_typing import TaskStatus
from ..log import (
get_logger,
get_console_log,
)
from .._daemon import maybe_spawn_daemon
if TYPE_CHECKING:
from ._sharedmem import (
ShmArray,
)
from .feed import _FeedsBus
log = get_logger(__name__)
# highest frequency sample step is 1 second by default, though in
# the future we may want to support shorter periods or a dynamic style
# tick-event stream.
_default_delay_s: float = 1.0
class Sampler:
'''
Global sampling engine registry.
Manages state for sampling events, shm incrementing and
sample period logic.
This non-instantiated type is meant to be a singleton within
a `samplerd` actor-service spawned once by the user wishing to
time-step sample real-time quote feeds, see
``._daemon.maybe_open_samplerd()`` and the below
``register_with_sampler()``.
'''
service_nursery: None | trio.Nursery = None
# TODO: we could stick these in a composed type to avoid
# angering the "i hate module scoped variables crowd" (yawn).
ohlcv_shms: dict[float, list[ShmArray]] = {}
# holds one-task-per-sample-period tasks which are spawned as-needed by
# data feed requests with a given detected time step usually from
# history loading.
incr_task_cs: trio.CancelScope | None = None
# holds all the ``tractor.Context`` remote subscriptions for
# a particular sample period increment event: all subscribers are
# notified on a step.
# subscribers: dict[int, list[tractor.MsgStream]] = {}
subscribers: defaultdict[
float,
list[
float,
set[tractor.MsgStream]
],
] = defaultdict(
lambda: [
round(time.time()),
set(),
]
)
@classmethod
async def increment_ohlc_buffer(
self,
period_s: float,
task_status: TaskStatus[trio.CancelScope] = trio.TASK_STATUS_IGNORED,
):
'''
Task which inserts new bars into the provide shared memory array
every ``period_s`` seconds.
This task fulfills 2 purposes:
- it takes the subscribed set of shm arrays and increments them
on a common time period
- broadcast of this increment "signal" message to other actor
subscribers
Note that if **no** actor has initiated this task then **none** of
the underlying buffers will actually be incremented.
'''
# TODO: right now we'll spin printing bars if the last time stamp is
# before a large period of no market activity. Likely the best way
# to solve this is to make this task aware of the instrument's
# tradable hours?
total_s: float = 0 # total seconds counted
ad = period_s - 0.001 # compensate for trio processing time
with trio.CancelScope() as cs:
# register this time period step as active
task_status.started(cs)
# sample step loop:
# includes broadcasting to all connected consumers on every
# new sample step as well incrementing any registered
# buffers by registered sample period.
while True:
await trio.sleep(ad)
total_s += period_s
# increment all subscribed shm arrays
# TODO:
# - this in ``numba``
# - just lookup shms for this step instead of iterating?
i_epoch = round(time.time())
broadcasted: set[float] = set()
# print(f'epoch: {i_epoch} -> REGISTRY {self.ohlcv_shms}')
for shm_period_s, shms in self.ohlcv_shms.items():
# short-circuit on any not-ready because slower sample
# rate consuming shm buffers.
if total_s % shm_period_s != 0:
# print(f'skipping `{shm_period_s}s` sample update')
continue
# update last epoch stamp for this period group
if shm_period_s not in broadcasted:
sub_pair = self.subscribers[shm_period_s]
sub_pair[0] = i_epoch
broadcasted.add(shm_period_s)
# TODO: ``numba`` this!
for shm in shms:
# print(f'UPDATE {shm_period_s}s STEP for {shm.token}')
# append new entry to buffer thus "incrementing"
# the bar
array = shm.array
last = array[-1:][shm._write_fields].copy()
# guard against startup backfilling races where
# the buffer has not yet been filled.
if not last.size:
continue
(t, close) = last[0][[
'time',
'close',
]]
next_t = t + shm_period_s
if shm_period_s <= 1:
next_t = i_epoch
# this copies non-std fields (eg. vwap) from the
# last datum
last[[
'time',
'open',
'high',
'low',
'close',
'volume',
]][0] = (
# epoch timestamp
next_t,
# OHLC
close,
close,
close,
close,
0, # vlm
)
# TODO: in theory we could make this faster by
# copying the "last" readable value into the
# underlying larger buffer's next value and then
# incrementing the counter instead of using
# ``.push()``?
# write to the buffer
shm.push(last)
# broadcast increment msg to all updated subs per period
for shm_period_s in broadcasted:
await self.broadcast(
period_s=shm_period_s,
time_stamp=i_epoch,
)
@classmethod
async def broadcast(
self,
period_s: float,
time_stamp: float | None = None,
) -> None:
'''
Broadcast the period size and last index step value to all
subscribers for a given sample period.
'''
pair = self.subscribers[period_s]
last_ts, subs = pair
task = trio.lowlevel.current_task()
log.debug(
f'SUBS {self.subscribers}\n'
f'PAIR {pair}\n'
f'TASK: {task}: {id(task)}\n'
f'broadcasting {period_s} -> {last_ts}\n'
# f'consumers: {subs}'
)
borked: set[tractor.MsgStream] = set()
for stream in subs:
try:
await stream.send({
'index': time_stamp or last_ts,
'period': period_s,
})
except (
trio.BrokenResourceError,
trio.ClosedResourceError
):
log.error(
f'{stream._ctx.chan.uid} dropped connection'
)
borked.add(stream)
for stream in borked:
try:
subs.remove(stream)
except ValueError:
log.warning(
f'{stream._ctx.chan.uid} sub already removed!?'
)
@classmethod
async def broadcast_all(self) -> None:
for period_s in self.subscribers:
await self.broadcast(period_s)
@tractor.context
async def register_with_sampler(
ctx: tractor.Context,
period_s: float,
shms_by_period: dict[float, dict] | None = None,
open_index_stream: bool = True, # open a 2way stream for sample step msgs?
sub_for_broadcasts: bool = True, # sampler side to send step updates?
) -> None:
get_console_log(tractor.current_actor().loglevel)
incr_was_started: bool = False
try:
async with maybe_open_nursery(
Sampler.service_nursery
) as service_nursery:
# init startup, create (actor-)local service nursery and start
# increment task
Sampler.service_nursery = service_nursery
# always ensure a period subs entry exists
last_ts, subs = Sampler.subscribers[float(period_s)]
async with trio.Lock():
if Sampler.incr_task_cs is None:
Sampler.incr_task_cs = await service_nursery.start(
Sampler.increment_ohlc_buffer,
1.,
)
incr_was_started = True
# insert the base 1s period (for OHLC style sampling) into
# the increment buffer set to update and shift every second.
if shms_by_period is not None:
from ._sharedmem import (
attach_shm_array,
_Token,
)
for period in shms_by_period:
# load and register shm handles
shm_token_msg = shms_by_period[period]
shm = attach_shm_array(
_Token.from_msg(shm_token_msg),
readonly=False,
)
shms_by_period[period] = shm
Sampler.ohlcv_shms.setdefault(period, []).append(shm)
assert Sampler.ohlcv_shms
# unblock caller
await ctx.started(set(Sampler.ohlcv_shms.keys()))
if open_index_stream:
try:
async with ctx.open_stream() as stream:
if sub_for_broadcasts:
subs.add(stream)
# except broadcast requests from the subscriber
async for msg in stream:
if msg == 'broadcast_all':
await Sampler.broadcast_all()
finally:
if sub_for_broadcasts:
subs.remove(stream)
else:
# if no shms are passed in we just wait until cancelled
# by caller.
await trio.sleep_forever()
finally:
# TODO: why tf isn't this working?
if shms_by_period is not None:
for period, shm in shms_by_period.items():
Sampler.ohlcv_shms[period].remove(shm)
if incr_was_started:
Sampler.incr_task_cs.cancel()
Sampler.incr_task_cs = None
async def spawn_samplerd(
loglevel: str | None = None,
**extra_tractor_kwargs
) -> bool:
'''
Daemon-side service task: start a sampling daemon for common step
update and increment count write and stream broadcasting.
'''
from piker._daemon import Services
dname = 'samplerd'
log.info(f'Spawning `{dname}`')
# singleton lock creation of ``samplerd`` since we only ever want
# one daemon per ``pikerd`` proc tree.
# TODO: make this built-into the service api?
async with Services.locks[dname + '_singleton']:
if dname not in Services.service_tasks:
portal = await Services.actor_n.start_actor(
dname,
enable_modules=[
'piker.data._sampling',
],
loglevel=loglevel,
debug_mode=Services.debug_mode, # set by pikerd flag
**extra_tractor_kwargs
)
await Services.start_service_task(
dname,
portal,
register_with_sampler,
period_s=1,
sub_for_broadcasts=False,
)
return True
return False
@acm
async def maybe_open_samplerd(
loglevel: str | None = None,
**kwargs,
) -> tractor._portal.Portal: # noqa
'''
Client-side helper to maybe startup the ``samplerd`` service
under the ``pikerd`` tree.
'''
dname = 'samplerd'
async with maybe_spawn_daemon(
dname,
service_task_target=spawn_samplerd,
spawn_args={'loglevel': loglevel},
loglevel=loglevel,
**kwargs,
) as portal:
yield portal
@acm
async def open_sample_stream(
period_s: float,
shms_by_period: dict[float, dict] | None = None,
open_index_stream: bool = True,
sub_for_broadcasts: bool = True,
cache_key: str | None = None,
allow_new_sampler: bool = True,
) -> AsyncIterator[dict[str, float]]:
'''
Subscribe to OHLC sampling "step" events: when the time aggregation
period increments, this event stream emits an index event.
This is a client-side endpoint that does all the work of ensuring
the `samplerd` actor is up and that mult-consumer-tasks are given
a broadcast stream when possible.
'''
# TODO: wrap this manager with the following to make it cached
# per client-multitasks entry.
# maybe_open_context(
# acm_func=partial(
# portal.open_context,
# register_with_sampler,
# ),
# key=cache_key or period_s,
# )
# if cache_hit:
# # add a new broadcast subscription for the quote stream
# # if this feed is likely already in use
# async with istream.subscribe() as bistream:
# yield bistream
# else:
async with (
# XXX: this should be singleton on a host,
# a lone broker-daemon per provider should be
# created for all practical purposes
maybe_open_samplerd() as portal,
portal.open_context(
register_with_sampler,
**{
'period_s': period_s,
'shms_by_period': shms_by_period,
'open_index_stream': open_index_stream,
'sub_for_broadcasts': sub_for_broadcasts,
},
) as (ctx, first)
):
async with (
ctx.open_stream() as istream,
# TODO: we don't need this task-bcasting right?
# istream.subscribe() as istream,
):
yield istream
async def sample_and_broadcast(
bus: _FeedsBus, # noqa
rt_shm: ShmArray,
hist_shm: ShmArray,
quote_stream: trio.abc.ReceiveChannel,
brokername: str,
sum_tick_vlm: bool = True,
) -> None:
'''
`brokerd`-side task which writes latest datum sampled data.
This task is meant to run in the same actor (mem space) as the
`brokerd` real-time quote feed which is being sampled to
a ``ShmArray`` buffer.
'''
log.info("Started shared mem bar writer")
overruns = Counter()
# iterate stream delivered by broker
async for quotes in quote_stream:
# print(quotes)
# TODO: ``numba`` this!
for broker_symbol, quote in quotes.items():
# TODO: in theory you can send the IPC msg *before* writing
# to the sharedmem array to decrease latency, however, that
# will require at least some way to prevent task switching
# at the yield such that the array write isn't delayed while
# another consumer is serviced..
# start writing the shm buffer with appropriate
# trade data
# TODO: we should probably not write every single
# value to an OHLC sample stream XD
# for a tick stream sure.. but this is excessive..
ticks = quote['ticks']
for tick in ticks:
ticktype = tick['type']
# write trade events to shm last OHLC sample
if ticktype in ('trade', 'utrade'):
last = tick['price']
# more compact inline-way to do this assignment
# to both buffers?
for shm in [rt_shm, hist_shm]:
# update last entry
# benchmarked in the 4-5 us range
o, high, low, v = shm.array[-1][
['open', 'high', 'low', 'volume']
]
new_v = tick.get('size', 0)
if v == 0 and new_v:
# no trades for this bar yet so the open
# is also the close/last trade price
o = last
if sum_tick_vlm:
volume = v + new_v
else:
# presume backend takes care of summing
# it's own vlm
volume = quote['volume']
shm.array[[
'open',
'high',
'low',
'close',
'bar_wap', # can be optionally provided
'volume',
]][-1] = (
o,
max(high, last),
min(low, last),
last,
quote.get('bar_wap', 0),
volume,
)
# TODO: PUT THIS IN A ``_FeedsBus.broadcast()`` method!
# XXX: we need to be very cautious here that no
# context-channel is left lingering which doesn't have
# a far end receiver actor-task. In such a case you can
# end up triggering backpressure which which will
# eventually block this producer end of the feed and
# thus other consumers still attached.
sub_key: str = broker_symbol.lower()
subs: list[
tuple[
tractor.MsgStream | trio.MemorySendChannel,
float | None, # tick throttle in Hz
]
] = bus.get_subs(sub_key)
# NOTE: by default the broker backend doesn't append
# it's own "name" into the fqsn schema (but maybe it
# should?) so we have to manually generate the correct
# key here.
fqsn = f'{broker_symbol}.{brokername}'
lags: int = 0
for (stream, tick_throttle) in subs.copy():
try:
with trio.move_on_after(0.2) as cs:
if tick_throttle:
# this is a send mem chan that likely
# pushes to the ``uniform_rate_send()`` below.
try:
stream.send_nowait(
(fqsn, quote)
)
except trio.WouldBlock:
overruns[sub_key] += 1
ctx = stream._ctx
chan = ctx.chan
log.warning(
f'Feed OVERRUN {sub_key}'
'@{bus.brokername} -> \n'
f'feed @ {chan.uid}\n'
f'throttle = {tick_throttle} Hz'
)
if overruns[sub_key] > 6:
# TODO: should we check for the
# context being cancelled? this
# could happen but the
# channel-ipc-pipe is still up.
if (
not chan.connected()
or ctx._cancel_called
):
log.warning(
'Dropping broken consumer:\n'
f'{sub_key}:'
f'{ctx.cid}@{chan.uid}'
)
await stream.aclose()
raise trio.BrokenResourceError
else:
await stream.send(
{fqsn: quote}
)
if cs.cancelled_caught:
lags += 1
if lags > 10:
await tractor.breakpoint()
except (
trio.BrokenResourceError,
trio.ClosedResourceError,
trio.EndOfChannel,
):
ctx = stream._ctx
chan = ctx.chan
if ctx:
log.warning(
'Dropped `brokerd`-quotes-feed connection:\n'
f'{broker_symbol}:'
f'{ctx.cid}@{chan.uid}'
)
if tick_throttle:
assert stream._closed
# XXX: do we need to deregister here
# if it's done in the fee bus code?
# so far seems like no since this should all
# be single-threaded. Doing it anyway though
# since there seems to be some kinda race..
bus.remove_subs(
sub_key,
{(stream, tick_throttle)},
)
# a working tick-type-classes template
_tick_groups = {
'clears': {'trade', 'dark_trade', 'last'},
'bids': {'bid', 'bsize'},
'asks': {'ask', 'asize'},
}
def frame_ticks(
first_quote: dict,
last_quote: dict,
ticks_by_type: dict,
) -> None:
# append quotes since last iteration into the last quote's
# tick array/buffer.
ticks = last_quote.get('ticks')
# TODO: once we decide to get fancy really we should
# have a shared mem tick buffer that is just
# continually filled and the UI just ready from it
# at it's display rate.
if ticks:
# TODO: do we need this any more or can we just
# expect the receiver to unwind the below
# `ticks_by_type: dict`?
# => undwinding would potentially require a
# `dict[str, set | list]` instead with an
# included `'types' field which is an (ordered)
# set of tick type fields in the order which
# types arrived?
first_quote['ticks'].extend(ticks)
# XXX: build a tick-by-type table of lists
# of tick messages. This allows for less
# iteration on the receiver side by allowing for
# a single "latest tick event" look up by
# indexing the last entry in each sub-list.
# tbt = {
# 'types': ['bid', 'asize', 'last', .. '<type_n>'],
# 'bid': [tick0, tick1, tick2, .., tickn],
# 'asize': [tick0, tick1, tick2, .., tickn],
# 'last': [tick0, tick1, tick2, .., tickn],
# ...
# '<type_n>': [tick0, tick1, tick2, .., tickn],
# }
# append in reverse FIFO order for in-order iteration on
# receiver side.
for tick in ticks:
ttype = tick['type']
ticks_by_type[ttype].append(tick)
# TODO: a less naive throttler, here's some snippets:
# token bucket by njs:
# https://gist.github.com/njsmith/7ea44ec07e901cb78ebe1dd8dd846cb9
async def uniform_rate_send(
rate: float,
quote_stream: trio.abc.ReceiveChannel,
stream: tractor.MsgStream,
task_status: TaskStatus = trio.TASK_STATUS_IGNORED,
) -> None:
# try not to error-out on overruns of the subscribed (chart) client
stream._ctx._backpressure = True
# TODO: compute the approx overhead latency per cycle
left_to_sleep = throttle_period = 1/rate - 0.000616
# send cycle state
first_quote = last_quote = None
last_send = time.time()
diff = 0
task_status.started()
ticks_by_type: defaultdict[
str,
list[dict],
] = defaultdict(list)
clear_types = _tick_groups['clears']
while True:
# compute the remaining time to sleep for this throttled cycle
left_to_sleep = throttle_period - diff
if left_to_sleep > 0:
with trio.move_on_after(left_to_sleep) as cs:
try:
sym, last_quote = await quote_stream.receive()
except trio.EndOfChannel:
log.exception(f"feed for {stream} ended?")
break
diff = time.time() - last_send
if not first_quote:
first_quote = last_quote
# first_quote['tbt'] = ticks_by_type
if (throttle_period - diff) > 0:
# received a quote but the send cycle period hasn't yet
# expired we aren't supposed to send yet so append
# to the tick frame.
frame_ticks(
first_quote,
last_quote,
ticks_by_type,
)
# send cycle isn't due yet so continue waiting
continue
if cs.cancelled_caught:
# 2 cases:
# no quote has arrived yet this cycle so wait for
# the next one.
if not first_quote:
# if no last quote was received since the last send
# cycle **AND** if we timed out waiting for a most
# recent quote **but** the throttle cycle is now due to
# be sent -> we want to immediately send the next
# received quote ASAP.
sym, first_quote = await quote_stream.receive()
frame_ticks(
first_quote,
first_quote,
ticks_by_type,
)
# we have a quote already so send it now.
with trio.move_on_after(throttle_period) as cs:
while (
not set(ticks_by_type).intersection(clear_types)
):
try:
sym, last_quote = await quote_stream.receive()
except trio.EndOfChannel:
log.exception(f"feed for {stream} ended?")
break
frame_ticks(
first_quote,
last_quote,
ticks_by_type,
)
# measured_rate = 1 / (time.time() - last_send)
# log.info(
# f'`{sym}` throttled send hz: {round(measured_rate, ndigits=1)}'
# )
first_quote['tbt'] = ticks_by_type
# TODO: now if only we could sync this to the display
# rate timing exactly lul
try:
await stream.send({sym: first_quote})
except (
# NOTE: any of these can be raised by ``tractor``'s IPC
# transport-layer and we want to be highly resilient
# to consumers which crash or lose network connection.
# I.e. we **DO NOT** want to crash and propagate up to
# ``pikerd`` these kinds of errors!
trio.ClosedResourceError,
trio.BrokenResourceError,
ConnectionResetError,
):
# if the feed consumer goes down then drop
# out of this rate limiter
log.warning(f'{stream} closed')
await stream.aclose()
return
# reset send cycle state
first_quote = last_quote = None
diff = 0
last_send = time.time()
ticks_by_type.clear()