Such that we maintain that subsys in the actor-runtime repo (with
hopefully an extensive test suite XD).
Port deats,
- rewrite `open_service_mngr()` as a thin wrapper that delegates into
the new `tractor.hilevel.open_service_mngr()` but with maintenance of
the `Services` class-singleton for now.
- port `.service._daemon` usage to the new
`ServiceMngr.start_service_ctx()` a rename from
`.start_service_task()` which is now likely destined for the soon
supported `tractor.trionics.TaskMngr` nursery extension.
- ref the new `ServiceMngr.an: ActorNursery` instance var name.
Other,
- always enable the `tractor.pause_from_sync()` support via `greenback`
whenever `debug_mode` is set at `pikerd` init.
Nothing changing functionally here just adding more `tractor`
operational notes, tips for debug tooling and typing fixes B)
Of particular note is adding further details about the reason we do not
need to call `Context.cancel()` inside the `finally:` block of
`.open_context_in_task()` thanks to `tractor`'s new and improved
inter-actor cancellation semantics Bo
Since `tractor`'s new and improved inter-actor cancellation semantics
are much more pedantic, AND bc we use the `ServiceMngr` for spawning
service actors on-demand, the caller of `maybe_spawn_daemon()` should
NEVER conduct a so called "out of band" `Actor`-runtime cancel request
since this is precisely the job of our `ServiceMngr` XD
Add a super in depth note explaining the underlying issue and adding
a todo list of how we should prolly augment `tractor` to make such cases
easier to grok and fix in the future!
Given it's a fairly simple yet useful abstraction, it makes sense to
offer this sub-sys alongside the core `tractor` runtime lib.
Without going into extreme detail on the impl changes (it'll come in
the commit that moves to the other repo) here is the high level summary:
------ - ------
- rename `Services` -> `ServiceMngr` and use an factory `@acm`
to guarantee a single-instance-per-actor using a niche approach for a
singleton object using a default keyword-arg B)
- the mod level `open_service_mngr()` and `get_service_mngr()` are the
new allocation/access API.
- add a `ServiceMngr.start_service()` method which does the work of both
spawning a new subactor (for the daemon) and uses its portal to start
the mngr side supervision task.
- open actor/task nurseries inside the `@acm` allocator.
Adjust other dependent subsystems to match:
------ - ------
- use `open_service_mngr()` when first allocated in `open_pikerd()`.
- use `get_service_mngr()` instead of importing the class ref inside
`.service.maybe_spawn_daemon()`, `.brokers._daemon.spawn_brokerd()`
and `.data._sampling.spawn_samplerd()` using a `partial` to pack in
the endpoint ctx kwargs (unpacked inside `.start_service()` XD).
A new class var `tuple[Exception]` such that the err set can be reffed
externally as needed for catching other similar pub-sub/IPC failures in
other (related) real-time sub-systems.
Also added some now-masked logging for debugging live-feed stream reading
issues that should ONLY be used for debugging since they'll greatly
degrade HFT perf. Used the new `log.mk_repr()` stuff (that one day we
should prolly pull from `modden` as a dep) for pretty console emissions.
Since now `tractor` will raise this native `trio`-exc translated from
a `Stop` msg when the peer gracefully terminates a `tractor.MsgStream`.
Just `info()` log in such cases versus continuing to warn for the
others.
Since currently we're only using this IPC subsys for `deribit`, and
generally speaking we're primarly supporting options markets (which are
fairly "slow moving"), flip to a default of NOT resetting the `NoBsWs`
on timeout since doing so normally breaks the jsron-rpc IPC session.
Without a proper `fixture` passed to `open_autorecon_ws()` (which we
should eventually implement!!) relying on a timeout-to-reset more or
less will just cause breakage issues - a proper reconnect sequence must
be implemented before using that feature.
Deats,
- expose and proxy through the `msg_recv_timeout` from
`open_jsonrpc_session()` into the underlying `open_autorecon_ws()`
call.
Add a doc-string reflecting recent refinements, drop all the old hook
params, rename `n: trio.Nursery` -> `tn` for "task nursery" fitting with
code base's naming style.
Namely handling backends which do not provide a default "frame
size-duration" in their init-config by making the backfiller guess the
value based on the first frame received.
Deats,
- adjust `start_backfill()` to take a more explicit
`def_frame_duration: Duration` expected to be unpacked from any
backend hist init-config by the `tsdb_backfill()` caller which now
also computes a value from the first received frame when the config
section isn't provided.
- in `start_backfill()` we now always expect the `def_frame_duration`
input and always decrement the query range by this value whenever
a `NoData` is raised by the provider-backend paired with an explicit
`log.warning()` about the handling.
- also relay any `DataUnavailable.args[0]` message from the provider
in the handler.
- repair "gap reporting" which checks for expected frame duration vs.
that received with much better humanized logging on the missing
segment using `pendulum.Interval/Duration.in_words()` output.
The `open_history_client()` provider endpoint can *optionally*
deliver a `frame_types: dict[int, pendulum.Duration]` subsection in its
`config: dict[str, dict]` (as was implemented with the `ib` backend).
This allows the `tsp` backfilling machinery to use this "recommended
frame duration" to subtract from the `last_start_dt` any time a `NoData`
gap is signalled by the `get_hist()` call allowing gaps to be ignored
safely without missing history by knowing the next earliest dt we can
query from using the `end_dt`. However, currently all crypto$ providers
haven't implemented this feat yet..
As such only try to use the `frame_types` feature if provided when
handling `NoData` conditions inside `tsp.start_backfill()` and otherwise
raise as normal.
Was orig for debugging an issue with `kucoin` i think but definitely
shouldn't be left in XD
Also add `'perpetual_future'` to the `.start_backfill()` input literal
set since we don't expect the 'btc/usd.perp.binance' for now.
Based on the default provided in their
[docs](https://docs.astral.sh/ruff/configuration/) and migrating
previous config from the prior `poetry`-verion of our `pyproject.toml`
Such that we start encouraging devs to lint code they touch and
hopefully we include a pass as part of our tests/CI eventually B)
Also, mk local `tractor` install `--editable` since without it being
a locally hackable repo it's kinda pointless to install from the local
fs Xp
Seems to be some kinda super weird env bug since we moved to using
`uv`? When it triggers it also seems to cause a pretty fundamental crash
that not only breaks `tractor.devx._debug` stuff but also seems to get
us in a perma-hang state where no SIGINT or other sys sig will be able
to kill the root proc!?!?
TODO, a `gitea` issue to track so we can fix the fundamental problem as
well as transitive fault in `tractor`'s core which seems to be due to
the error taking place during a sub-actor's module import phase which
prevents the runtime from booting fully and then the proc getting stuck
in a real gnarly SIG-state..
This must have broke at some point during the new `MktPair` and thus
`.fqme: str` updates; mas-o-menos the symbol key in the quote-msg-`dict`
was NOT set to the `MktPair.bs_fqme: str` value and thus wasn't being
processed by the downstream sampling and feed subsys.
So fix that as well as a few other refinements,
- set the `topic: mkt.bs_fqme` in quote msgs obvi.
- drop the "wait for first clearing vlm" quote poll loop; going to fix
the sampler to handle a `first_quote` without a `'last'` key.
- add some typing around calls to `get_mkt_info()`.
- rename `stream_messages()` -> `iter_normed_quotes()`.
In line with our move to `uv` and recent `nix` support update a bunch of
the summary content and factor out the order-control section to a new
`.piker.clearing` readme file with embedded todos therein.
Like other backends use the `AsyncClient` for all venue specific
client-sessions but change to allocating them inside `get_client()`
using an `AsyncExitStack` and inserting directly in the
`Client.venue_sesh: dict` table during init.
Supporting impl tweaks:
- remove most of the API client session building logic and instead make
`Client.__init__()` take in a `venue_sessions: dict` (set it to
`.venue_sesh`) and `conf: dict`, instead opting to do the http client
configuration inside `get_client()` since all that code only needs to
be run once.
|_load config inside `get_client()` once.
|_move session token creation into a new util func `init_api_keys()` and
also call it from `get_client()` factory; toss in an ex. toml section
config to the doc string.
- define `_venue_urls: dict[str, str]` (content taken from old static
`.venue_sesh` dict) at module level and feed them as `base_url: str`
inputs to the client create loop.
- adjust all call sigs in httpx-sesh-using methods, namely just
`._api()`.
- do a `.exch_info()` call in `get_client()` to cache the symbology
set.
Unrelated changes for various other outstanding buggers:
- to get futures feeds correctly loading when selected
from search (like 'XMRUSDT.USDTM.PERP'), expect a `MktPair` input to
`Client.bars()` such that the exact venue-key can be looked up (via
a new `.pair2venuekey()` meth) and then passed to `._api()`.
- adjust `.broker.open_trade_dialog()` to failover to paper engine when
there's no `api_key` key set for the `subconf` venue-key.
Like we do with other history backends to indicate lack of a data set.
This avoids any raise that will will bring down the backloader task with
some downstream error.
Raise a `ValueError` on no time index for now.
About time we tidy'd a buncha status logging in this backend..
particularly for boot-up where there's lots of client-try-connect poll
looping with account detection from the user config.
`.api.Client` pprint and logging fmt improvements:
- add `Client.__repr__()` which shows the minimally useful set of info
from the underlying `.ib: IB` as well as a new `.acnts: list[str]`
of the account aliases defined in the user's `brokers.toml`.
- mk `.bars()` define a comprehensive `query_info: str` with all the
request deats but only display if there's a problem with the response
data.
- mk `.get_config()` report both the config file path and the acnt
aliases (NOT the actual account #s).
- move all `.load_aio_clients()` client poll loop requests do
`log.runtime()` statuses, only falling through to a `.warning()` when
the loop fails to connect the client to the spec-ed API-gw addr, and
|_ don't allow loading accounts for which the user has not defined an
alias in `brokers.toml::[ib]`; raise a value-error in such cases
with a message indicating how to mod the config.
|_ only `log.info()` about acnts if some were loaded..
Other mod logging de-noising:
- better status fmting in `.symbols.open_symbol_search()` with
`repr(Client)`.
- for `.feed.stream_quotes()` first quote reporting use `.runtime()`.
Tyler Goodlet
Nelson Torres
goodboy