We're doing this in `Flow.update_graphics()` atm and probably are going
to in general want custom graphics objects for all the diff curve / path
types. The new flows work seems to fix the bounding rect width calcs to
not require the ad-hoc extra `+ 1` in the step mode case; before it was
always a bit hacky anyway. This also tries to add a more correct
bounding rect adjustment for the `._last_line` segment.
Finally this gets us much closer to a generic incremental update system
for graphics wherein the input array diffing, pre-graphical format data
processing, downsampler activation and incremental update and storage of
any of these data flow stages can be managed in one modular sub-system
:surfer_boi:.
Dirty deatz:
- reorg and move all path logic into `Renderer.render()` and have it
take in pretty much the same flags as the old
`FastAppendCurve.update_from_array()` and instead storing all update
state vars (even copies of the downsampler related ones) on the
renderer instance:
- new state vars: `._last_uppx, ._in_ds, ._vr, ._avr`
- `.render()` input bools: `new_sample_rate, should_redraw,
should_ds, showing_src_data`
- add a hack-around for passing in incremental update data (for now)
via a `input_data: tuple` of numpy arrays
- a default `uppx: float = 1`
- add new render interface attrs:
- `.format_xy()` which takes in the source data array and produces out
x, y arrays (and maybe a `connect` array) that can be passed to
`.draw_path()` (the default for this is just to slice out the index
and `array_key: str` columns from the input struct array),
- `.draw_path()` which takes in the x, y, connect arrays and generates
a `QPainterPath`
- `.fast_path`, for "appendable" updates like there was on the fast
append curve
- move redraw (aka `.clear()` calls) into `.draw_path()` and trigger
via `redraw: bool` flag.
- our graphics objects no longer set their own `.path` state, it's done
by the `Flow.update_graphics()` method using output from
`Renderer.render()` (and it's state if necessary)
A bit hacky to get all graphics types working but this is hopefully the
first step toward moving all the generic update logic into `Renderer`
types which can be themselves managed more compactly and cached per
uppx-m4 level.
Which is basically just "deleting" rows from a column series.
You can only use the trim command from the `.cmd` cli and only with a so
called `LocalClient` currently; it's also sketchy af and caused
a machine to hang due to mem usage..
Ideally we can patch in this functionality for use by the rpc api
and have it not hang like this XD
Pertains to https://github.com/alpacahq/marketstore/issues/264
Yet another path ops routine which converts a 1d array into a data
format suitable for rendering a "step curve" graphics path (aka a "bar
graph" but implemented as a continuous line).
Also, factor the `BarItems` rendering logic (which determines whether to
render the literal bars lines or a downsampled curve) into a routine
`render_baritems()` until we figure out the right abstraction layer for
it.
Starts a module for grouping together all our `QPainterpath` related
generation and data format operations for creation of fast curve
graphics. To start, drops `FastAppendCurve.downsample()` and moves
it to a new `._pathops.xy_downsample()`.
Mostly just dropping old commented code for "step mode" format
generation. Always slice the tail part of the input data and move to the
new `ms_threshold` in the `pg` profiler'
Relates to the bug discovered in #310, this should avoid out-of-order
msgs which do not have a `.reqid` set to be error logged to console.
Further, add `pformat()` to kraken logging of ems msging.
Since downsampling with the more correct version of m4 (uppx driven
windows sizing) is super fast now we don't need to avoid downsampling
on low uppx values. Further all graphics objects now support in-view
slicing so make sure to use it on interaction updates. Pass in the view
profiler to update method calls for more detailed measuring.
Even moar,
- Add a manual call to `.maybe_downsample_graphics()` inside the mouse
wheel event handler since it seems that sometimes trailing events get
lost from the `.sigRangeChangedManually` signal which can result in
"non-downsampled-enough" graphics on chart given the scroll amount;
this manual call seems to entirely fix this?
- drop "max zoom" guard since internals now support (near) infinite
scroll out to graphics becoming a single pixel column line XD
- add back in commented xrange signal connect code for easy testing to
verify against range updates not happening without it
This took longer then i care to admit XD but it definitely adds a huge
speedup and with only a few outstanding correctness bugs:
- panning from left to right causes strange trailing artifacts in the
flows fsp (vlm) sub-plot but only when some data is off-screen on the
left but doesn't appear to be an issue if we keep the `._set_yrange()`
handler hooked up to the `.sigXRangeChanged` signal (but we aren't
going to because this makes panning way slower). i've got a feeling
this is a bug todo with the device coordinate cache stuff and we may
need to report to Qt core?
- factoring out the step curve logic from
`FastAppendCurve.update_from_array()` (un)fortunately required some
logic branch uncoupling but also meant we needed special input controls
to avoid things like redraws and curve appends for special cases,
this will hopefully all be better rectified in code when the core of
this method is moved into a renderer type/implementation.
- the `tina_vwap` fsp curve now somehow causes hangs when doing erratic
scrolling on downsampled graphics data. i have no idea why or how but
disabling it makes the issue go away (ui will literally just freeze
and gobble CPU on a `.paint()` call until you ctrl-c the hell out of
it). my guess is that something in the logic for standard line curves
and appends on large data sets is the issue?
Code related changes/hacks:
- drop use of `step_path_arrays_from_1d()`, it was always a bit hacky
(being based on `pyqtgraph` internals) and was generally hard to
understand since it returns 1d data instead of the more expected (N,2)
array of "step levels"; instead this is now implemented (uglily) in
the `Flow.update_graphics()` block for step curves (which will
obviously get cleaned up and factored elsewhere).
- add a bunch of new flags to the update method on the fast append
curve: `draw_last: bool`, `slice_to_head: int`, `do_append: bool`,
`should_redraw: bool` which are all controls to aid with previously
mentioned issues specific to getting step curve updates working
correctly.
- add a ton of commented tinkering related code (that we may end up
using) to both the flow and append curve methods that was written as
part of the effort to get this all working.
- implement all step curve updating inline in `Flow.update_graphics()`
including prepend and append logic for pre-graphics incremental step
data maintenance and in-view slicing as well as "last step" graphics
updating.
Obviously clean up commits coming stat B)
Since we have in-view style rendering working for all curve types
(finally) we can avoid the guard for low uppx levels and without losing
interaction speed. Further don't delay the profiler so that the nested
method calls correctly report upward - which wasn't working likely due
to some kinda GC collection related issue.
More or less this improves update latency like mad. Only draw data in
view and avoid full path regen as much as possible within a given
(down)sampling setting. We now support append path updates with in-view
data and the *SPECIAL CAVEAT* is that we avoid redrawing the whole curve
**only when** we calc an `append_length <= 1` **even if the view range
changed**. XXX: this should change in the future probably such that the
caller graphics update code can pass a flag which says whether or not to
do a full redraw based on it knowing where it's an interaction based
view-range change or a flow update change which doesn't require a full
path re-render.
After much effort (and exhaustion) but failure to get a view into our
`numpy` OHLC struct-array, this instead allocates an in-thread-memory
array which is updated with flattened data every flow update cycle.
I need to report what I think is a bug to `numpy` core about the whole
view thing not working but, more or less this gets the same behaviour
and minimizes work to flatten the sampled data for line-graphics drawing
thus improving refresh latency when drawing large downsampled curves.
Update the OHLC ds curve with view aware data sliced out from the
pre-allocated and incrementally updated data (we had to add a last index
var `._iflat` to track appends - this should be moved into a renderer
eventually?).
This begins the removal of data processing / analysis methods from the
chart widget and instead moving them to our new `Flow` API (in the new
module introduce here) and delegating the old chart methods to the
respective internal flow. Most importantly is no longer storing the
"last read" of an array from shm in an internal chart table (was
`._arrays`) and instead the `ShmArray` instance is passed as input and
stored in the `Flow` instance. This greatly simplifies lookup logic such
that the display loop now doesn't have to worry about reading shm, it
can be done by internal graphics logic as desired. Generally speaking,
all previous `._arrays`/`._graphics` lookups are now delegated to the
entries in the chart's `._flows` table.
The new `Flow` methods are generally better factored and provide more
detailed output regarding data-stream <-> graphics inter-relations for
the future purpose of allowing much more efficient update calls in the
display loop as well as supporting low latency interaction UX.
The concept here is that we're introducing an intermediary layer that
ties together graphics and real-time data flows such that widget code is
oriented around plot layout and the flow apis are oriented around
real-time low latency updates and providing an efficient high level
metric layer for the UX.
The summary api transition is something like:
- `update_graphics_from_array()` -> `.update_graphics_from_flow()`
- `.bars_range()` -> `Flow.datums_range()`
- `.bars_range()` -> `Flow.datums_range()`
Given that naming the port map is mostly pointless, since accounts can
be detected once the client connects, just expect a `brokers.toml` to
define a simple sequence of port numbers. Toss in a warning for using
the old map/`dict` style.
Now that we have working client auth thanks to:
https://github.com/barneygale/asyncvnc/pull/4 and related issue,
we can use a pw for the vnc server, though we should eventually
auto-generate a random one from a docker super obviously.
Add logic to the data reset hack loop to do a connection reset after
2 failed/timeout attempts at the regular data reset. We need to also add
this logic around reconnectionn events that are due to the host
network connection: aka roaming that's faster then timing logic
builtin to the gateway.
`ib-gw` seems particularly fragile to connections from clients with the
same id (can result in weird connect hangs and even crashes) and
`ib_insync` doesn't handle intermittent tcp disconnects that
well..(especially on dockerized IBC setups). This adds a bunch of
changes to our client caching and scan loop as well a proper
task-locking-to-cache-proxies so that,
- `asyncio`-side clients aren't double-loaded/connected even when
explicitly trying to reconnect repeatedly with a given client to work
around the unreliability of the `asyncio.Transport` design in
`ib_insync`.
- we can use `tractor.trionics.maybe_open_context()` to lock the `trio`
side from loading more then one `Client` on the `asyncio` side and
instead on cache hits only making a new `MethodProxy` around the
reused `asyncio`-side client (since each `trio` task needs its own
inter-task msg channel).
- a `finally:` block teardown on all clients loaded in the scan loop
avoids stale connections.
- the connect params are now exposed as named args to
`load_aio_clients()` can be easily controlled from caller code.
Oh, and we properly hooked up the internal `ib_insync` logging to our
own internal schema - makes it a lot easier to debug wtf is going on XD
In order to expose more `asyncio` powered `Client` methods to endpoint
task-code this adds a more extensive and layered set of `MethodProxy`
loading routines, in dependency order these are:
- `load_clients_for_trio()` a `tractor.to_asyncio.open_channel_from()`
entry-point factory for loading all scanned clients on the `asyncio` side
and delivering them over the inter-task channel to a `trio`-side task.
- `get_preferred_data_client()` a simple client instance loading routine
which reads from the users `brokers.toml -> `prefer_data_account:
list[str]` which must list account names, in priority order, that are
acceptable to be used as the main "data connection client" such that
only one of the detected clients is used for data (whereas the rest
are used only for order entry).
- `open_client_proxies()` which delivers the detected `Client` set
wrapped each in a `MethodProxy`.
- `open_data_client()` which directly delivers the preferred data client
as a proxy for `trio` tasks.
- update `open_client_method_proxy()` and `open_client_proxy` to require
an input `Client` instance.
Further impl details:
- add `MethodProxy._aio_ns` to ref the original `asyncio` side proxied instance
- add `Client.trades()` to pull executions from the last day/session
- load proxies inside `trades_dialogue` and use the new `.trades()`
method to try and pull a fill ledger for eventual correct pp price
calcs (pertains to #307)..
We return a copy (since since a view doesn't seem to work..) of the
(field filtered) shm array contents which is the same index-length as
the source data.
Further, fence off the resource tracker disable-hack into a helper
routine.
It seems once in a while a frame can get missed or dropped (at least
with binance?) so in those cases, when the request erlangs is already at
max, we just manually request the missing frame and presume things will
work out XD
Further, discard out of order frames that are "from the future" that
somehow end up in the async queue once in a while? Not sure why this
happens but it seems thus far just discarding them is nbd.
Bleh/🤦, the ``end_dt`` in scope is not the "earliest" frame's
`end_dt` in the async response queue.. Parse the queue's latest epoch
and use **that** to compare to the last last pushed datetime index..
Add more detailed logging to help debug any (un)expected datetime index
gaps.
When the tsdb has a last datum that is in the past less then a "frame's
worth" of sample steps we need to slice out only the data from the
latest frame that doesn't overlap; this fixes that slice logic..
Previously i dunno wth it was doing..
When the market isn't open the feed layer won't create a subscriber
entry in the sampler broadcast loop and so if a manual call to
``broadcast()`` is made (like when trying to update a chart from
a history prepend) we need to handle that case and just broadcast
a random `-1` for now..BD
Expect each backend to deliver a `config: dict[str, Any]` which provides
concurrency controls to `trimeter`'s batch task scheduler such that
backends can define their own concurrency limits.
The dirty deats in this patch include handling history "gaps" where
a query returns a history-frame-result which spans more then the typical
frame size (in seconds). In such cases we reset the target frame index
(datetime index sequence implemented with a `pendulum.Period`) using
a generator protocol `.send()` such that the sequence can be dynamically
re-indexed starting at the new (possibly) pre-gap datetime. The new gap
logic also allows us to detect out of order frames easier and thus wait
for the next-in-order to arrive before making more requests.
Use the new `open_history_client()` endpoint/API and expect backends to
provide a history "getter" routine that can be called to load historical
data into shm even when **not** using a tsdb. Add logic for filling in
data from the tsdb once the backend has provided data up to the last
recorded in the db. Add logic for avoiding overruns of the shm buffer
with more-then-necessary queries of tsdb data.
It turns out (i guess not so shockingly?) that `marketstore` doesn't
always teardown "gracefully" under SIGINT (seems to hang if there are
open client connections which are also in the midst of teardown?) so
this instead first tries the SIGINT and then fails over to a SIGKILL
(destroy loop) which seems to be much more reliable to ensure shutdown
without any downside - in terms of a "hard kill".
Originally i was thinking the issue was root perms related (which get
relegated solely to the `marketstored` daemon actor after spawn) but
actually it was indeed the signalling / application layer causing the
hold-up/latency on teardown. There's a bunch of lingering (now
commented) code which tried to solve this non-problem as well as a bunch
logging/prints to help decipher the root of the issue - this will all
get cleaned out shortly.
If `marketstore` is detected try to only load most recent missing data
from the data provider (broker) and the rest from the tsdb and push it
all to shm for display in the UI. If the provider/broker doesn't have
the history client endpoint, just use the old one for now so we can
start to incrementally add support. Don't start the ohlc step
incrementer task until the backend signals that the feed is live.
Add some basic `numpy` epoch slice logic to generate append and prepend
arrays to write to the db.
Mooar cool things,
- add a `Storage.delete_ts()` method to wipe a column series from the db
easily.
- don't attempt to read in any OHLC series by default on client load
- add some `pyqtgraph` profiling and drop manual latency measures
- if no db series for the fqsn exists write the entire shm array
Also, Start tinkering with `tractor.trionics.ipython_embed()`
In effort to get back to a usable REPL around the mkts client
this adds usage of the new `tractor` integration api as well as logic
for skipping backfilling if existing tsdb arrays are found.
Starts a wrapper around the `marketstore` client to do basic ohlcv query
and retrieval and prototypes out write methods for ohlc and tick.
Try to connect to `marketstore` automatically (which will fail if not
started currently) but we will eventually first do a service query.
Further:
- get `pikerd` working with and without `--tsdb` flag.
- support spawning `brokerd` with no real-time quotes.
- bring back in "fqsn" support that was originally not
in this history before commits factoring.
Not sure how I missed mapping the 5995 grpc port 🤦; done now.
Also adds graceful teardown using SIGINT with included container
logging relayed to the piker console B).
Found this caused breakage on `kraken` orders which triggered the
"insufficient funds" error response. Makes sense since they won't
generate an order id if the order can't ever be submitted.
The most important changes include:
- iterating the new `Flow` type and updating graphics
- adding detailed profiling
- increasing the min uppx before graphics updates are throttled
- including the L1 spread in y-range calcs so that you never have the
bid/ask go "out of view"..
- pass around `Flow`s instead of shms
- drop all the old prototyped downsampling code
If manually managing an overlay you'll likely call `.overlay_plotitem()`
and then a plotting method so we need to accept a plot item input so
that the chart's pi doesn't get assigned incorrectly in the `Flow` entry
(though it is by default if no input is provided).
More,
- add a `Flow.graphics` field and set it to the `pg.GraphicsObject`.
- make `Flow.maxmin()` return `None` in the "can't calculate" cases.
- set shm refs on `Flow` entries.
- don't run a graphics cycle on 'update' msgs from the engine
if the containing chart is hidden.
- drop `volume` from flows map and disable auto-yranging
once $vlm comes up.
Allows for removing resize callbacks for a flow/overlay that you wish to
remove from view (eg. unit volume after dollar volume is up) and thus
less general interaction callback overhead for any plot you don't wish
to show or resize.
Further,
- drop the `autoscale_linked_plots` block for now since with
multi-view-box overlays each register their own vb resize slots
- pull the graphics object from the chart's `Flow` map inside
`.maybe_downsample_graphics()`
This new type wraps a shm data flow and will eventually include things
like incremental path-graphics updates and serialization + bg downsampling
techniques. The main immediate motivation was to get a cached y-range max/min
calc going since profiling revealed the `numpy` equivalents were
actually quite slow as the data set grows large. Likely we can use all
this to drive a streaming mx/mn routine that's always launched as part
of each on-host flow.
This is our official foray into use of `msgspec.Struct` B) and I have to
say, pretty impressed; we'll likely completely ditch `pydantic` from
here on out.
We don't need update graphics on every x-range change since that's what
the display loop does. Instead, only on manual changes do we make manual
calls into `.update_graphics_from_array()` and be sure to iterate all
linked subplots and all their embedded graphics.
The pg profiler seems to have trouble with early `return`s in function
calls (likely muckery with the GC/`.__delete__()`) so let's just try
to avoid it for now until we either fix it (probably by implementing as
a ctx mngr) or use diff one.
Ugh, turns out the wacky `ChartView.maxmin` callback stuff we did (for
determining y-range sizings) currently requires that the volume array
has a "bars in view" result.. so let's make that keep working without
rendering the graphics for the curve (since we're disabling them once
$vlm comes up).
As with the `BarItems` graphics, this makes it possible to pass in a "in
view" range of array data that can be *only* rendered improving
performance for large(r) data sets. All the other normal behaviour is
kept (i.e a persistent, (pre/ap)pendable path can still be maintained)
if a ``view_range`` is not provided.
Further updates,
- drop the `.should_ds_or_redraw()` and `.maybe_downsample()` predicates
instead moving all that logic inside `.update_from_array()`.
- disable the "cache flipping", which doesn't seem to be needed to avoid
artifacts any more?
- handle all redraw/dowsampling logic in `.update_from_array()`.
- even more profiling.
- drop path `.reserve()` stuff until we better figure out how it's
supposed to work.
Drop all the logic originally in `.update_ds_line()` which is now done
internal to our `FastAppendCurve`. Add incremental update of the
flattened OHLC -> line curve (unfortunately using `np.concatenate()` for
the moment) and maintain a new `._ds_line_xy` arrays tuple which keeps
the internal state. Add `.maybe_downsample()` as per the new interaction
update method requirement. Draft out some fast path curve stuff like in
our line graphic. Short-circuit bars path updates when we downsample to
line. Oh, and add a ton more profiling in prep for getting
all this stuff faf.
Build out an interface that makes it super easy to downsample curves
using the m4 algorithm while keeping our incremental `QPainterPath`
update feature. A lot of hard work and tinkering went into getting this
working all in-thread correctly and there are quite a few details..
New interface methods:
- `.x_uppx()` which returns the x-axis "view units per pixel"
- `.px_width()` which returns the total (rounded) x-axis pixels spanned
by the curve in view.
- `.should_ds_or_redraw()` a predicate which checks internal state to
see if either downsampling of the curve should take place, or the curve
should have all downsampling removed and be redrawn with source array
data.
- `.downsample()` the actual ds processing routine which delegates into
the m4 algo impl.
- `.maybe_downsample()` a simple update method which can be called by
the view box when the user changes the zoom level.
Implementation details/changes:
- make `.update_from_array()` check for downsample (or revert to source
aka de-downsample) conditions exist and then downsample and re-draw
path graphics accordingly.
- in order to even further speed up path appends (since our main
bottleneck is measured to be `QPainter.drawPath()` calls with large
paths which are frequently updates), add a secondary path `.fast_path`
which is the path that is real-time updates by incremental appends and
which is painted separately for speed in `.pain()`.
- drop all the `QPolyLine` stuff since it was tested to be much slower
in general and especially so for append-updates.
- stop disabling the cache settings on updates since it doesn't seem to
be required any more?
- more move toward deprecating and removing all lingering interface
requirements from `pg.PlotCurveItem` (like `.xData`/`.yData`).
- adjust `.paint()` and `.boundingRect()` to compensate for the new
`.fast_path`
- add a butt-load of profiling B)
Tyler Goodlet
dinkus
Guillermo Rodriguez