开云体育

• I received 451,009 TickByTickLast objects for NQZ3
• It took 445 requests/responses and 1,045 seconds to receive all data
• that is an average of 2.347s per request (elapsed time)
• but the median was sub-second at 0.659s.

• it is a single threaded event processor (no sleeps, delays, or built-in pacing)
• requests data in 1,000 tick chunks in reverse time order
• converts each returned TickByTickLast into a relatively expensive immutable Java object
• accumulates the objects in a list that is serialized into streamable Json objects and written to file each time more than 10,000 ticks have been accumulated

That means the tool makes about 9 out of 10 requests immediately (a few micro seconds) after the callback for the previous request. There is a short processing delay before every tenth request (5ms to 40ms) for the data serialization and file storage.

Now, the interesting finding is the discrepancy between average and median response times. IBKR paced the responses within chunks of ~60 seconds, each time with roughly the following rhythm:

• Ten requests with response times around 600ms each
• One request with 3.5 seconds
• Three requests with 600ms each
• One request with 4.5 seconds
• One request with 600ms
• One request with 5.5 seconds
• Three requests with 6 seconds

• One request with 12 seconds

Attached a couple charts of what that looked like. I do not have data on how long they keep that 60 second chunk rhythm up in case you download a couple years worth of data. My gut tells me that you will not be able to keep up the less-than 3 second average for very long runs.

闯ü谤驳别苍

First of all, order status ApiPending is not an error condition. It is the state of your order(s) right after you placed them. And the section of the does provide high-level descriptions of when to expect the various order stats. They say for :

ApiPending - Indicates order has not yet been sent to IB server, for instance if there is a delay in receiving the security definition. Uncommonly received.

I'd review the contract definition and you should seriously think about requesting a well configured contract from IBKR via instead of initializing various fields yourselves. For example, I am not sure why you set a "strike" for am STK instrument, and initialize trading class, symbol, and local symbol will have important fields initialized with meaningful values.

And then you use a Time in Force of IOT. Does the behavior change if you place the order with a simpler TIF of, say GTC or DAY?

闯ü谤驳别苍

I'm now having a seemingly identical problem using IB's python client -- I can only imagine I'm doing something boneheaded, but without documentation on the error, I'm a bit stuck.

Running on v177, I set up an order like so:

??????? con = Contract()
??????? con.conId = 76792991
??????? con.symbol = "TSLA"
??????? con.secType = "STK"
??????? con.strike = 0.
??????? con.exchange = "SMART"
??????? con.primaryExchange = "NASDAQ"
??????? con.currency = "USD"
??????? con.tradingClass = "NMS"
??????? con.localSymbol = "TSLA"

??????? order = Order()
??????? order.action = "BUY"
??????? order.totalQuantity = 1.
??????? order.orderType = "MKT"
??????? order.tif = "IOC"
??????? order.account = "[account name]"
??????? app.placeOrder([order id], con, order)

This sends:

3-[order id]-76792991-TSLA-STK--0.0---SMART-NASDAQ-USD-TSLA-NMS---BUY-1.0-MKT---IOC--[account name]--0--1-0-0-0-0-0-0-0--0-------0---1-0---0---0-0--0------0-----0-----------0---0-0---0--0-0-0-0--1.7976931348623157e+308-1.7976931348623157e+308-1.7976931348623157e+308-1.7976931348623157e+308-1.7976931348623157e+308-0----1.7976931348623157e+308-----0-0-0--2147483647-2147483647-0---

... but a call to openOrders gives

OrderStatus. Id: [order id] Status: ApiPending Filled: 0 Remaining: 1 AvgFillPrice: 0 PermId: 0 ParentId: 0 LastFillPrice: 0 ClientId: 3 WhyHeld:? MktCapPrice: 0

As I say, I imagine I'm setting something up incorrectly. If anyone can point me towards what (or even where to look), I'd be super grateful.

On Sun, Oct 1, 2023 at 03:15 PM, David Armour wrote:

I agree with everything you said although I am not entirely happy with my final solution. I do not feel that destroying the EReader object is the correct way to disconnect. It is certainly not intuitive which means I cannot be the only person finding this problem. Is everybody using C++ just ignoring this error?

You may see this as a wart but I, personally, won't comment with regard to how intuitive things are or any design aesthetic. I will say that consideration should be given to backward compatibility and just how much existing code would break (and therefore need to change) if the disconnect where removed from the destructor.

Just some food for thought... since there's usually ripple effects, unintended consequences, and API stability is actually very important. Also, I don't think it's wise to hold off fixing some simple, obvious, and existing issue merely because there's an intention of addressing it in a more significant re-design. As they say, there is no time like the present and I appreciate small and tractable, incremental changes.

However, this may come down to a judgment call. Since, I guess, people have indeed been ignoring it or working around it (or it's simply gone unnoticed). After all, it only happens late in the game when there usually isn't much network activity expected or necessarily happening.

Anyway, I will sleep on it one more night before submitting a bug report.

Is there a special place to report API bugs or do we just raise ticket as usual?

A pull request is probably the best way to go about this:

On Sun, Oct 1, 2023 at 10:02 AM, David Armour wrote:

Thanks for your comment Buddy.

You're welcome. I think you've found an interesting oversight.

I agree that the logic would seem that setting m_isAlive to false during the ~EReader() call should prevent the code trying to read from the socket.

Yeah, superficially and at first blush it would seem that way. But it'd be a mistake to think so. I can understand how that initial thought may have fooled the original author into thinking the code was correct though. All this is post-mortem speculation on my part however.

Nevertheless, once considered more thoughtfully, it becomes apparent that nothing prevents the destructor from being called while the processing thread's readToQueue is in the while block.

The std::atomic will only insure that reading and writing m_isAlive from multiple threads is well-defined; but guarantees nothing w.r.t. the destructor and what's happening in the while loop.

I tried playing around with that. If I pause after m_isAlive is set to false, then continue, the code works successfully without 509 error. It is as if at

"Pause", with a sleep or something? No... that's surely a game of whack-a-mole. It might work sometimes, depending on unpredictable factors, and might not.

It can be fun to experiment with the timing while you investigate though. You can also try playing w/ ordering by running the code on a single core/cpu and see how that affects things.

full speed the atomic operation is not being completed before the call to eDisconnect() happens. On Visual Studio this even happens when I turn off all Optimisations. As eDisconnect() does not use anything related to that atomic, I wonder if the compiler is parallelising the operation as all atomic operations are very slow.

Compiler optimizations fall somewhat into the category of "unpredictable factors" I alluded to (although they'd be deterministic given the compiler source and enough work). It's a bit surprising that a critical section wasn't implemented via mutex, but maybe the author wanted to avoid a performance penalty.

I also noticed that eDisconnect() sets m_fd = -1.? This should prevent the call to processNonBlockingSelect() from trying to call onReceive() which ultimately triggers receive() and recv() causing the error. Again the setting of the atomic variable is not being done in time.

I am still scratching my head over this. I suspect some synchronization issue between the various threads.

Anyway... I think you found a legitimate problem and I wouldn't worry about it too much. I like your proposed solution as well. Joining the thread and then disconnecting makes clear sense to me. And, as long as m_pClientSocket isn't being shared all over the place I don't see a reason to use more complicated locking via mutex.

If I were the reviewer I could see signing off after some confirmation tests... open a ticket and see what they have to say. If you submit a patch remember to change the IB_POSIX ifdef block too :-)

I guess the assumption is that using std::atomic<bool> m_isAlive; turns the following into a critical section:

void EReader::readToQueue() {
	//EMessage *msg = 0;

	while (m_isAlive) {
		if (m_buf.size() == 0 && !processNonBlockingSelect() && m_pClientSocket->isSocketOK())
			continue;

        if (!putMessageToQueue())
			break;
	}

But it does not, so I take your point.

??? if (m_hReadThread) {
??? ? ? ?? m_isAlive = false;
?? ??? ??? m_pClientSocket->eDisconnect();
? ?? ????? WaitForSingleObject(m_hReadThread, INFINITE);
?? ???? }

??? if (m_hReadThread) {
????????? m_isAlive = false;
? ? ????? WaitForSingleObject(m_hReadThread, INFINITE);
?? ?????? m_pClientSocket->eDisconnect();
?? ???? }

????? ftrMsgProcThrd_ = pool_->submit(
????????? [&]()
????????? {
??????????? while (clientSocket_->isConnected())
??????????? {
????????????? signal_.waitForSignal();?? // This waits 2 seconds.
????????????? reader_->processMsgs();
??????????? }
????????? });

I decided to tackle this annoying bug (not the first time) and have found that after the call to EClientSocket::eDisconnect() which calls EClientSocket::SocketClose() which just calls a Windows Sockets closesocket() on the open socket, I am still getting the message processing thread (EReader thread) trying to perform a Windows Sockets recv() messages on the closed socket resulting in a 509 error. I have traced that error to be socket error 10038 which confirms it is an invalid socket (in this case, a closed socket).

Before the line "reader_->processMsgs()"? I have tried checking for the socket still being open with if(clientSocket_->isConnected()) but it does not solve the problem. The EReader is running in its own thread as per the reader_->start() call.

I thought, perhaps I need to close the EReader before calling eDisconnect so I tried deleting the object and removing the call to eDisconnect() because the destructor of the EReader calls eDisconnect() itself but this does not fix the error. i still get the 509 caused by a read on the closed socket.

I am struggling here and would appreciate advise from any C++ coders that use a multi-threaded approach like the above. It is likely a threading issue but if anyone else has faced a similar "disconnect" issue I would be happy to hear what you did to resolve it.

"200ms-400ms? delay in between each order is transmitted by Gateway and has?Submitted status"

• A laptop on which I just did a fresh reinstall of Windows
• A newly-downloaded TWS (version 10.25.1j, which is their 'LATEST' build, but I've also tried the most recent STABLE and BETA builds, same result)
• A newly-downloaded TWS API version 10.25.01?(which is LATEST, but also tested STABLE, which is v 10.19.01, same result)
• And the Excel file that generates the NBBO error is the default 'LegacyTwsDde.xls' file that gets downloaded along with the API. I literally just open it in its default state, put my username is cell D5 of the Basic Orders sheet, and press the 'Place / Modify Order' button on one of the default populated rows (e.g. IBM, row 15)...and it throws the NbboPriceCap error in cell J6, as shown in my screenshot.

@Richard

I know that you were not trying to suggest one should use the same numerical ids for different request types. I just wanted to add another reason why one should not do it.

Looks like we "late comers" have a little advantage since the TWS API source code contains more useful bits than what you had to work with in the early years.

@Gordon

We are meandering away from the original topic, but a couple thoughts on why you get "117" or something similar for nextValidId.

TWS/IBGW actually memorizes the highest orderId for each clientId and each account. I did not know this either for the longest time, but I think it was JG who made a comment related to this in a post a couple months ago. So I went through the logs and looked for the oldest clientId I could find for a client that placed orders (that clientId had not been used in at least three years), used that clientId to connect to the account, and for sure, automagically, nextOrderId upon connection returned a number that was one higher than the last order that client had placed three years ago.

So if you are reusing the same clientId and that client occasionally places orders, nextOrderId will be going up slightly over time. I guess the clientId that gave you a nextValidId of "117" has placed 115 or so orders over time.

You can break that cycle and reset all orderIds for all clientIds for an account back to, I believe, "1" with the "Reset API order ID sequence" button at the bottom of "Global Configuration -> Configuration -> API -> Settings".

We also never call reqIds while a client is running. The nextValidId we receive as part of the connection protocol is sufficient for a sequence of (thread safe) incrementing orderIds.

We recently added a little code that potentially nudges the clients' internal nextValidOrderId counter while they are running. That code assures that the nextValidOrderId counter is always higher than any orderId the client is exposed to by openOrder and orderStatus callbacks. This is most important for the master client and client 0 but we just added it to the framework for all clients based on this comment in :

"However if there are multiple client applications connected to one account, it is necessary to use an order ID with new orders which is greater than all previous order IDs returned to the client application in openOrder or orderStatus callbacks."

闯ü谤驳别苍

I never got a first nextValidID? < 100.?
Got things like "117", why ? never look at the why, I don't know of any zombie process and it works, so ...

Your absolutely right, rules are more relax, and I am a bad guy, I don't abide all rules that I may recommend. However I was burned being too relax.

You have more experience than I on handling IB and have the practical knowledge of the limits in use. As your approach work and fulfill your need, and your right, classifying request is not a must.
To be a better guy ... hereafter sharing what I do. (IMHO, just a use case) not the right thread for that but the djinn is out of the bottle

1- I use a single central requestId 'pump', thread safe,
and another one, separate, for OrderId.
I used to use the same pump for both, but it's not pretty. (very subjective criteria);

2- I always track nextValidId and use it as seed for it.
Taking the highest of next one or IB suggestion.
Handling of nextValidID must be within the protected critical section of the TS pump.
But it's very rare, happens only on connect.

3- I don't call reqIds , not that it's bad, just that I am not ready to wait for the answer.

4- I prefer "overpumping" (wasting ids) than taking the risk to re-use an existing id, even if serviced (even if it may work, I feel it's like playing with fire)

5- I try to keep continuity in OrderId, (I do in fact, but the code is ready for any discontinuity)
But I don't care at all about continuity on 'request'.
Assuming that the higher is the reqid the latest the request was built, was enough for diag.

The side effect is that OrderId grow much slower than reqid, so rapidly you can diagnose what you are looking at just by the range of the number.
@Jurgen, de facto it does something like what you suggest with +100000, (which make more sense, using a fixed value, I agree, as its very unlikely you ever do 100000 orders without a reboot)
However again human-wise: it's difficult to read a huge number.
Most of the time, tests are conducted at beginning of code start, hence small numbers which are easier to memorize or even simply read.

6- I did try to make it bit field on LSB,
In fact in was not so handy as sorting pending/send/received request id manually during debug can become cumbersome.
It help tracking the class of request but you loose the ease to track the sequence.

I dislike ranges, because I always wonder if or when they may overlap. (or require code to handle it as an exception) Not enough experiences with that, probably too dogmatic

This is just a use case.