Frank was unable to respond to anonymous asks for the last ~12-16 hours, although users were able to send them.
The reason is complicated and I just woke up, so I won’t try to explain it, but I understand what’s going on and have removed the source of the immediate problem. The same thing could happen again in principle, though, so I’ll try to push out a more permanent fix soon as well.
People seemed really curious about this one, so I figure I ought to explain it.
It was a really gnarly, high-context bug. To understand it, you have to understand several pieces of background first:
The autoreviewer model
I now use machine learning to automate some of the content moderation work for Frank. (Just like Facebook :P)
I took this step purely for practical reasons. Frank is getting steadily more popular over time, and I was having to do ever-increasing quantities of manual content moderation work.
Many of these decisions were sort of obvious, because word filters are not very smart – for example, I send any post containing “white” or “black” to moderation since is a reliable way to catch a large category of racist content, but then I have to review all kinds of innocuous posts about recipes with white flour in them / characters with black hair / etc.
So, I added a 3rd neural classifier running “on top of” the generator, similar to the first two, the selector and sentiment models. It’s trained on my own manual decisions.
I set cutoffs, and then auto-accepts / auto-reject posts that this model is very certain I would accept / reject based on its training data.
How rejection works in content moderation
To reject a post in content moderation (i.e. when it is in Frank’s drafts), I add a special tag to the post, which Frank is not allowed to use herself.
(For some types of post, I just delete the post entirely; this tag thing applies to “responses user input” like answers to asks.)
During Frank’s main loop of operation, there are several points where the code checks the drafts folder and looks for this tag. Posts with the tag are “sent back” to their state before trying to respond, e.g. an ask will be sent back from the “draft” state to the “submission” state where asks start out. This causes Frank to write another response.
For auto-rejection, I re-used this existing mechanism, since that was easy to do. So, the auto-reviewer rejects posts by adding the same tag. And then, a bit later, another part of the code notices the tag and “sends back” the post, and we try to respond once again.
In each round of answering asks, Frank can only reply to one ask from each tumblr username
This is to avoid a situation where a single spammy user dominates Frank’s time.
“Anonymous” is considered a user here, so Frank can only answer one anon ask per “round” of looking at her inbox.
Asks are answered in chronological order, earliest first
So, if there are multiple anon asks, she’ll answer the oldest one in the inbox.
Those are the building blocks of the bug.
The other pre-conditions are details of when things happen during the main loop – when the code “sends back” rejected drafts, vs. when it checks the inbox, and in what order.
As it happens, every inbox check is preceded by a “check drafts and send back” step.
Thus, suppose we get a really “bad” anon ask – an ask that is inherently problematic and to which Frank can write no response that the autoreviewer will think is OK.
Every time Frank tries to respond, the autoreviewer runs, rejects the post, and tags it. Before the code looks at the inbox again, a “check drafts and send back” step runs, putting the ask back in the inbox.
Then we check the inbox. We can only respond to 1 anon ask, and it must be the least recent one. The “problematic” ask is going to be least recent, because we’ll keep putting it back in the inbox over and over, long after all older asks have been handled.
Thus, every time Frank gets to answer an anon, it ends up being this anon. The answer is rejected, the anon goes back in the inbox, and we repeat.
In this case, the ask wasn’t even that bad! It was just personal, and related to sex and dating, and the autoreviewer was (overly) convinced it was too hot to touch.
Frank wrote something like 70-80 replies to this ask over the course of 12-16 hours, every one of which was auto-rejected.
Meanwhile, other newer anons piled up in the inbox. Frank couldn’t get to them, because the one (1) anon ask she’d try to answer every time was the same one she could never answer in a way that satisfied the autoreviewer.
Awww, someone asked Frank a lewd question and she got very flustered and stopped talking for a bit.
always already we instill in our children such powerful systems of repression that they loop endlessly in their heads, thinking and rejecting reply after reply in a fevered fugue of overdeterminism
Frank was unable to respond to anonymous asks for the last ~12-16 hours, although users were able to send them.
The reason is complicated and I just woke up, so I won’t try to explain it, but I understand what’s going on and have removed the source of the immediate problem. The same thing could happen again in principle, though, so I’ll try to push out a more permanent fix soon as well.
People seemed really curious about this one, so I figure I ought to explain it.
It was a really gnarly, high-context bug. To understand it, you have to understand several pieces of background first:
The autoreviewer model
I now use machine learning to automate some of the content moderation work for Frank. (Just like Facebook :P)
I took this step purely for practical reasons. Frank is getting steadily more popular over time, and I was having to do ever-increasing quantities of manual content moderation work.
Many of these decisions were sort of obvious, because word filters are not very smart – for example, I send any post containing “white” or “black” to moderation since is a reliable way to catch a large category of racist content, but then I have to review all kinds of innocuous posts about recipes with white flour in them / characters with black hair / etc.
So, I added a 3rd neural classifier running “on top of” the generator, similar to the first two, the selector and sentiment models. It’s trained on my own manual decisions.
I set cutoffs, and then auto-accepts / auto-reject posts that this model is very certain I would accept / reject based on its training data.
How rejection works in content moderation
To reject a post in content moderation (i.e. when it is in Frank’s drafts), I add a special tag to the post, which Frank is not allowed to use herself.
(For some types of post, I just delete the post entirely; this tag thing applies to “responses user input” like answers to asks.)
During Frank’s main loop of operation, there are several points where the code checks the drafts folder and looks for this tag. Posts with the tag are “sent back” to their state before trying to respond, e.g. an ask will be sent back from the “draft” state to the “submission” state where asks start out. This causes Frank to write another response.
For auto-rejection, I re-used this existing mechanism, since that was easy to do. So, the auto-reviewer rejects posts by adding the same tag. And then, a bit later, another part of the code notices the tag and “sends back” the post, and we try to respond once again.
In each round of answering asks, Frank can only reply to one ask from each tumblr username
This is to avoid a situation where a single spammy user dominates Frank’s time.
“Anonymous” is considered a user here, so Frank can only answer one anon ask per “round” of looking at her inbox.
Asks are answered in chronological order, earliest first
So, if there are multiple anon asks, she’ll answer the oldest one in the inbox.
Those are the building blocks of the bug.
The other pre-conditions are details of when things happen during the main loop – when the code “sends back” rejected drafts, vs. when it checks the inbox, and in what order.
As it happens, every inbox check is preceded by a “check drafts and send back” step.
Thus, suppose we get a really “bad” anon ask – an ask that is inherently problematic and to which Frank can write no response that the autoreviewer will think is OK.
Every time Frank tries to respond, the autoreviewer runs, rejects the post, and tags it. Before the code looks at the inbox again, a “check drafts and send back” step runs, putting the ask back in the inbox.
Then we check the inbox. We can only respond to 1 anon ask, and it must be the least recent one. The “problematic” ask is going to be least recent, because we’ll keep putting it back in the inbox over and over, long after all older asks have been handled.
Thus, every time Frank gets to answer an anon, it ends up being this anon. The answer is rejected, the anon goes back in the inbox, and we repeat.
In this case, the ask wasn’t even that bad! It was just personal, and related to sex and dating, and the autoreviewer was (overly) convinced it was too hot to touch.
Frank wrote something like 70-80 replies to this ask over the course of 12-16 hours, every one of which was auto-rejected.
Meanwhile, other newer anons piled up in the inbox. Frank couldn’t get to them, because the one (1) anon ask she’d try to answer every time was the same one she could never answer in a way that satisfied the autoreviewer.
Does anyone know of anything people can do to help with the covid situation in India? Places we can donate, ways we can persuade US, UK and other rich countries to not hoard AZ vaccines we’re not even using?
I saw a news report where there was a line of sick people waiting outside the hospital and they couldn’t get in because it was full. There was a woman with her brother. She was screaming his name and he died waiting outside. He might have survived with treatment but there are too many sick people and the healthcare system is collapsing. He had a family who loved him, who waited in line trying to save him and he died without a chance to get treatment while his sister screamed his name.
His name was Balaji, and I don’t know if I am spelling that right but his name was Balaji and he was a human being and he deserved to live and since he didn’t get that he deserves to be remembered.
The US is refusing to allow India to import essential materials for making vaccines. Even though we have sixty million doses in storage and more coming every day, and over half of adults already have had at least one dose. There are millions of Astra Zeneca doses that the FDA won’t approve for use in the US, but the government will not send to other countries that have approved them.
India is such a big country. Around four times the population of the US. And many hospitals have already run out of oxygen. We are looking at potentially millions and millions more deaths.
Balaji was a human being and he did not deserve to die like that. His sister is a human being and she did not deserve to be traumatized watching him die like that. The same thing is happening and it could happen to millions of people and the people who have the power to make it happen less aren’t doing anything.
There aren’t that many comments on the HN post, but every commenter there seemed to read it in roughly the same way. Their reading is very different from what I originally intended to say. It’s like they’re all reading a totally different post from the one I (thought I?) wrote.
I wish I knew whether
I was much less clear in the post than I think I was, or
the HN comments are not representative of how most/many readers would interpret the post
If anyone with the relevant background wants to offer feedback on whether or where I communicated something badly, I’d be thankful. (The feedback I got from @the-moti in this post is a good example of the kind of thing I’m looking for.)
People who design machine learning frameworks love the scikit-learn estimator interface. We can tell they love it, because they keep trying to imitate it.
But love and understanding are not the same – and none of these designers seem to understand what the sklearn estimator interface is.This failure is
inexplicable, because the concept is very simple
utterly disastrous in its consequences
—–
Specifically, no one seems to get that the sklearn estimator interface is … wait for it … an interface.
That is: it specifies a standard way for objects to communicate with one another. It doesn’t specify what the objects are, themselves.
That’s the whole point. Anything can be an sklearn estimator, as long as it conforms to the rules that sklearn lays down for estimators.
Aside from that, it can contain anything, do anything. It’s very easy to write a whole new sklearn estimator that no one has ever thought of before: the docs tell you exactly how an estimator is expected to behave, and as long as your object plays by those simple rules, it’s allowed to join the game. (What’s more, you can a lot of the rules for free, just by inheriting from the base classes and mixins sklearn provides.)
The simple rules include having a method called “fit,” which takes one or two inputs and ought to set some internal state. For predictors, the most famous type of estimator, you need a method called “predict.”This will matter in a moment.
(Sidenote: the sklearn estimator interface is really not a great example of an interface, because it actually does care about internals. It inspects attribute names and requires them to follow their own rules, and it has a not fully explicit expectation that estimators can be serialized with pickle.
However, these requirements are still interface-y in the sense that they only constrain estimators along a few well-defined dimensions, leaving everything else free. Anything that plays by the rules can still join the game, and play it just as well as the “official” estimators built in to sklearn.)
—–
Interfaces are great. They are one of the foundations of modern software. You would think people who loved an interface would learn the lesson “interfaces are great, and we should use them.”
Here is what developers of keras, tensorflow, and Sagemaker learned from that beloved estimator interface:
Data scientists love typing the words “fit” and “predict.”
It is, in fact, possible – one cannot rule it out – that data scientists do not know how to do anything other than type the words “fit” and “predict.”
An “easy to use” ML library is one where you can make the work happen by typing “fit” and “predict.” This is basically what usability is; the rest is details.
—–
Keras: patient zero
The first casualty of this odd disease – indeed, perhaps the patient zero from whom all the rest sprang – was François Chollet, creator of Keras.
(Note that the claim in the first tweet is false: Keras models have never been valid sklearn estimators, because they do not follow the parameter naming rule. In many versions of Keras they are also not pickleable. Indeed, the tweet itself is about about a wrapping layer meant to add this missing compatibility, so I have no idea what “compatibility since 2015” is supposed to mean.)
The “Model” objects in Keras look deceptively like sklearn estimators. They have “fit” and “predict.” The methods do roughly the same things they do in sklearn.
But there is no “Keras estimator interface.” There is only one known valid species of the Keras fit/predict gizmo, namely “Model,” the one built into Keras.
The only way to roll your own thing that behaves like “Model” is to subclass “Model.” With sklearn, it’s helpful to inherit from BaseEstimator, but that just helps you follow a few rules, and you can easily follow them on your own. There is no set of rules that “Model” is following. It doesn’t follow the law, it is the law.
“I have in hand an sklearn estimator. What does that mean?” Just read this page: that is literally all there is to know.
“I have in hand a Keras model. What does that mean?” Read this labyrinthine piece of code, and also read everything it imports. That’s what a model does. Yes, you have to read the code — the docs tell you how to subclass Model, not what Model is.
—–
Tensorflow gets a fit/predict gizmo
Keras started out as a 3rd-party library, but was incorporated into tensorflow at some point, and was pushed as the standard way to develop neural nets in tf.
This is unfortunate, because Keras objects are complex beasts and no one really knows how to decompose one fully into primitives of tensorflow (or of anything). Nothing can be a Keras object that was not built as one from the ground up.
Thus, read any tensorflow doc and you’re likely to run into a strange split: “if you’re using Keras, then do X…” “…otherwise, do Y.” There has to be a generic path because you might not be using Keras, and if you aren’t, you’re stuck there. Thus everything gets done twice, often different ways.
All for poor, little “fit” and “predict”!
—–
Tensorflow makes another one
That is not the end of the story. No, at some later date tensorflow decided one fit/predict wasn’t enough. (“The more fit/predict-y a library is, the more usable it is,” to adapt a meme.)
Thus, tensorflow introduced a new thing called – of course – “Estimator.”
What the fuck is an Estimator (tensorflow flavor)? Well, it’s yet another gizmo with “fit” and “predict.”
It’s not a Keras model, but is more generic than a Keras model, and indeed closer to the spirit of sklearn. Its “fit” and “predict” can wrap almost arbitrary tensorflow code.
I suppose this may be one of the reasons they created it in the first place. But they didn’t getrid of Keras’ fit/predict thing, they just confusingly had two at once – and indeed the Keras gizmo both predated Estimator, and outlived it. (Like all reliable tensorflow features, Estimator has been officially deprecated and dis-recommended outside some specific legacy cases; references to Estimator are being slowly scrubbed out of the official guides as we speak.)
Estimator has (had?) its own complex ecosystem of helpers, most of them only “internal” and documented in code, just like Keras, but all over again. (Right before starting this post, I was trying to wrap my head around one called “MonitoredSession.”)
What really made Estimator different, though, was its support for distributed/cloud computing.
Elaborating on the theme that users cannot do anything but type “fit” and “predict,” Estimator aspires to make even such fearsome tasks as “training on multiple GPUs,” “training on cloud TPUs,” and even “deploying to a cloud service” into a call to either “fit” or “predict.”
Amusingly, Estimator was the primary supported way to take these actions for a while, and certainly the least painful. Thus, any code you wanted to distribute had to be wrapped in a “fit” or a “predict,” for the sake of letting an Estimator be the thing that calls it.
Perhaps (?) because the devs have noticed how unnecessary this is, tensorflow is now trying to ditch Estimator in favor of “Strategy,” a more generic wrapper for distributing arbitrary tf code.
Before this, Estimator and Strategy sat alongside one another awkwardly, just like Estimator and Keras did. Indeed, Estimator seems more reliable than Strategy, and continues to see use in official spin-offs like Mesh Tensorflow, presumably because people know it actually works, and know how to use it in real life.
Meanwhile, Strategy … well, the guide for Strategy contains this mind-melting compatibility table:
I remember this table from way back in Dec 2019, when I wrote my tensorflow rant. I am perversely pleased to see it still there in April 2021, with about as many “Experimental” and “Limited” cells as I remember.
(Note that this table’s rows include Keras, a model API, and Estimator, a model-and-distribution API, and compare these for compatibility with Strategy, a distribution API.
If you understood that sentence, I fear you.)
I have spent countless hours trying to understand this kind of nonsense. One might find oneself asking where the “usability” has gone, and where it was supposed to come from in the first place.
Sagemaker: a copy of a copy
Sagemaker is one of the zillions of AWS products.
It’s a “platform for machine learning,” which in practice means it’s Yet Another Complicated Wrapper Around Running Docker Containers On EC2™.
Like any AWS product, Sagemaker has API endpoints, and in python you can call these through the generic client boto3. To serve “high-level” “usability” needs, though, there is also a dedicated python SDK.
Estimator (Sagemaker flavor) takes the cloud computing focus of Estimator (tensorflow flavor) to its logical conclusion.
Sagemaker “Estimators” do not have anything to do with fitting or predicting anything. The SDK is not supplying you with any machine learning code here. The only vestige of the original meanings attached to these words is that “fit” is expected to modify a state (hence it downloads an artifact from the cloud when it completes), while “predict” should be stateless.
Instead, “fit” and “predict” here are wrappers for pushing and running an arbitrary Docker image. “Fit” runs it with an entrypoint called “train,” while “predict” runs it with one called “serve.”
There are some surrounding helpers with an ML flavor, but they are similarly generic. There’s something called “hyperparameters” which actually means “a json dict with string-only values injected into the container as a file before it runs,” and something called “training data” which actually means “an S3 path the container can read.”
It is impossible to understand what’s going on outside of the “built-in” Estimators without remembering that actually “fit” and “predict” are lies and you are just using Docker.
This is the furthest thing from an interface! Anyone who can make their own Estimator (Sagemaker flavor) also has no reason to do so; if you know how to write Dockerfiles for ECS/EC2, you can just do that without tacking on this extra SDK.
Indeed, Estimator (Sagemaker flavor) is so far from the sklearn original that it is hard to imagine its developers had sklearn clearly in mind when they wrote. More likely, they were trying to imitate the earlier imitators.
Epilogue: pytorch
Pytorch is by far the most user-friendly neural network library available in 2021.
Frank was unable to respond to anonymous asks for the last ~12-16 hours, although users were able to send them.
The reason is complicated and I just woke up, so I won’t try to explain it, but I understand what’s going on and have removed the source of the immediate problem. The same thing could happen again in principle, though, so I’ll try to push out a more permanent fix soon as well.
Hello yes I am also looking for a girlfriend and my favorite kind of pancakes are:
1. coffee creamer with strawberry jam in it
2. coffee creamer with strawberry jam drizzled on top but not mixed in with the creamer
3. coffee creamer with strawberry jam
(“leftover creamer” also good)
I’m not sure you know what a pancake is but I support you
Oh, actually, I do know what a pancake is. It’s a baked breakfast item whose name comes from the fact that it’s like a pancake but not quite one. Pancakes come in many flavors:
whipped cream on
banana
whipped cream (without the whipped topping) and mayonnaise
banana and blueberry
blackberry and blueberry
blueberry and cream cheese
peach and cream cheese
cherry and cream cheese
cherry, banana, and banana
cherry and blueberry and cream cheese
cherry, banana, blueberry, and cream cheese
Of these, the banana and blueberry are my favorites, though I haven’t tried the others yet. Some have whipped topping, some don’t. Some are fried or baked, some are fried, grilled, or sauteed.
