This requirement is convenient for the client but makes implementing robust servers harder than it need be. If instead each message was identified by a {timestamp,UUID} tuple then multiple MX servers could do final delivery to an eventually-consistent shared store. The requirement for strict, durable message ordering is the one thing that forces hard synchronization among multiple machines doing delivery.
POP3 actually made a step this direction with the common "UIDL" extension; IMAP4 felt like going backwards on this.
While I'm on the subject, another less-common POP3 extension I liked was "XTND XMIT". This gave a simple way for the POP3 client to send mail; sadly it never really caught on in favor of just using SMTP submission. Originally though this only worked OK because SMTP was open-relay. Once spam put an end to that it became complicated to send a message when accessing your mail remotely. Then to solve that problem we had to graft an auth layer on top of the existing SMTP protocol and upgrade all of the clients and servers to support it. If only "XTND XMIT" had won in the beginning everybody would be submitting outgoing mail over that already-authenticated channel and so much work could have been avoided.
So anyway, those are my two requests for any "IMAP killer": allow the server to use unique (but non-sequential) IDs for messages and provide a way to submit outgoing messages over the same channel.
But one thing that you MUST have for sensible synchronisation is the global MODSEQ. Otherwise there's no single state you can use to know if there's new data on the server short of (as you do with POP3) fetching the entire UIDL list every time.
The way we do this with Cyrus replication is to reinject the message if we get a UID clash - so you can have delivery at both ends of a (theoretical - not yet finished) master<=>master replica pair. Upon discovering the mismatch, BOTH messages get new UIDs larger than any yet used, which brings all clients into agreement.
But anyway - our unique identifiers in JMAP can be anything the server would like. We use 'fNUMBERuNUMBER' at FastMail, which is a folder/uid pair - but a more modern server could use UUIDs that don't change for moves between folders. We could even implement it in Cyrus on top of the Digest.SHA1 field if we built a reverse mapping database and deduplicated on it.
But synchronised modseqs... I don't see a way around that. We've been discussing "modseq validity range" stuff - where a delivery logs an intent to use a modseq, and then clears the intent upon commit. Any request that includes a range past that intent needs to wait for the commit to complete or abort.
1. Get everything created since last sync. A created index (as opposed to updated index) would not need to be resorted for subsequent updates. This would get only new mail.
2. A fixed-size Merkle tree, 8 blocks per level, 7 levels deep, with around 262144 blocks at the base. This would be the best tree configuration for syncing from a few thousand up to 40 million emails. This would identify all mail updated since last connection, within 4 roundtrips with a minimum of bandwidth. The tree should be incrementally updated using XOR at both client and server whenever an update is made. It's possible to use a tree with less depth for fewer roundtrips but with support for less emails or requiring more bandwidth.
3. Push sync. Server would push all creates/updates to client while client is connected.
Also, I strongly believe that per-message UUIDs should always start with the delivery time. That way as long as your servers are all reasonably synchronized you can get reasonable "sort by folder order" behavior without any sequential ID.
You're right that the modseq issue is hard, but the Merkle tree suggestion from @jorangreef can address that. Refer to my reply to him for more details.
Agreed on the messageid thing. I frankly think that IMAP is too messy to try too hard for interop with. Using a relatively similar data model and being able to work correctly with a maildir should be considered enough.
But I love this idea and hope it takes off in spite of that. It's really time to revamp the interface to mail servers to be more friendly to webmail. It'll make a huge difference to the level of choice of webmail providers/software.
A proper mail submission command would preserve parity with SMTP by keeping envelope and headers distinct. I just want to leverage the fact that I've already auth'ed the user and shouldn't have to redo that to a separate server.
Also, presumably one of the advantages to the REST+JSON design of JMAP is to make it easy for javascript clients to use it. Since they can't do direct SMTP anyway there will presumably be some REST service that they will use for submission.
It just seems that it should be part of the same standard. Most mail readers are also mail senders.
I don't understand this criticism, the criticism is usually that IMAP does not require maintaining UIDs (the server can announce UIDNOTSTICKY[0])
> While I'm on the subject, another less-common POP3 extension I liked was "XTND XMIT". This gave a simple way for the POP3 client to send mail
https://news.ycombinator.com/item?id=4826447
> IMAP has actually been extended to support that feature (not just kind of randomly, but by the IETF "Lemonade" working group), even allowing (in concert with other extensions) compositing a message from parts (such as attachments) on the server without having to download them first to the client.
> However, and I think this is actually more relevant in this context, Mark Crispin was an "opponent". (Note: the following e-mail snippet by Mark Crispin, was written years after the Lemonade Submit proposal expired, and thereby should be considered to already take into account that context.)
>> For many years, there have been various proposals to add mail sending capabilities to mail access protocols such as POP and IMAP.
>> These proposals are always strongly opposed. It is one of the "attractive nuisances" of email protocols. The value of the capability is obvious to many people, but the high cost of having it in POP or IMAP is much less obvious.
>> I am one of the opponents. For the past 25 or so years we have been in the overwhelming majority. It is quite unlikely that this concensus will change. If anything, it has become stronger in recent years.
>> […]
>> If you're not really that curious, suffice it to say that the people who design the protocols and systems understand the attraction but have excellent reasons not to do it.
http://mailman2.u.washington.edu/pipermail/imap-uw/2009-Janu...
[0] http://mailman2.u.washington.edu/pipermail/imap-protocol/200...
Their reasons aren't excellent, they are some theoretical notion of purity which leads to things like the "MOVE" capability taking years to finally happen, and everyone implementing copy+store+expunge independently and poorly in the meanwhile.
He's dead now, so he can't tell me I'm a feckless Gen X with no clue how to do real protocols - but there's a reason why proprietary protocols like ActiveSync get traction, and it's not because you need to pay a licence fee - it's because they actually solve problems that need solving.
We will be adding sending to the JMAP spec, although support may be optional. The way it currently works at FastMail is a bit too specific to our architecture though, so we want to clean it up before adding it to the spec.
Yes, exactly. UIDs should be sticky, but they should also be allocatable by widely distributed nodes without the need for central coordination.
If your UIDs are 32-bit monotonically-increasing integers then this is impossible. If they are 128-bit random numbers you get it for free. If you prefix them with a timestamp you even get reasonable ordering.
The entire UIDNOTSTICKY problem is a result of IMAP UIDs being so restricted.
I can see why a proposal as complicated as Lemonade failed to get traction, but I find the argument pretty thin. The "attractive nuisance" was the old open-relay SMTP infrastructure. If you were writing a POP client in 1991 you wouldn't think twice about sending mail since you just had to hit a SMTP server (didn't even have to be the "right" one!) and the mail would get delivered. When this all came crashing down far more effort got put into authenticated SMTP (talk about not separating concerns!) than would have been required to just get message submission right in the first place.
IMAP4rev1 is one thing, but there are many extensions supported in modern imap servers to speed up syncing and mailing, like CATENATE/BURL/CONDSTORE/MULTIAPPEND/MOVE/etc
Granted, the IMAP protocol is pretty hairy and difficult to work with. On the other hand, there's a huge ginormous deployment of it on the client and server, and the IETF WG behind it. I doubt JMAP will replace it anytime soon unless the WG itself takes up the issue. And given the neckbeards there ;-), some of whom have been working in IMAP for decades, you face the same opposition to change as those trying to legalize drugs or gay marriage. ;-)
We're still trying to get IPv6 deployed!
Our plan is to port our web interface across to this spec and implement both a direct backed in Cyrus IMAPd and a standalone implementation which can proxy for existing IMAP servers.
But I'm not sure it's a good idea to emphasize the fact that you use JSON. Sure, it's cool right now, but I find it technically more interesting that it replaces IMAP with something that is built on top of HTTP, a somewhat RESTful protocol.
(don't even get me started on trying to write a proxy for IMAP that's aware of the connection state)
Those things called standards generally find their way through bodies like the IETF. Maybe I missed it while reading on my phone, but I don't see this submitted as an IETF draft...
JMAP isn't the be-all and end-all of mail protocols, but it goes some way to address some of the problems in IMAP. We're using it in production right now at FastMail so we already know its pretty good as a low-latency, stateless protocol.
As for the IETF (and standards bodies in general), they're useful in their place, but they're also a slow-moving beast and thev're published a great many "standards" that have either never been implemented or are just rubbish.
But there's nobody forcing you to implement it. Its there, its documented, and we're encouraging others to get involved to make it work. If others get on board then we'll see improvements everywhere. If not, then FastMail will continue to use it and get the advantages that come from it.
Anyhow, this is all for a “Standard track” RFC. An “Informational” or “Experimental” RFC has no such requirement, and can basically be submitted by anyone for anything.
For me instant notification of inbound message is a must-have.
Our actual push does include some extra detail, but the important bit is just "here's a new state that exists on the server - if you haven't already heard of it, you probably should poll for changes now".
It's out of band(ish) - because JMAP is connectionless. And because depending on which platform you're using, there might be a nice channel for push notifications already.
In any case, I think you could have something a lot closer to REST, even if you carry on with grouping calls together.
Something like:
[
[ "GET", "/messages", { "search": "foo" }],
[ "GET", "/mailboxes", { "etag": "bar" }]
]
REST is way too slow. The JMAP format massively reduces round trip times, especially when you have sequential operations where you must wait for one to finish before the next can happen. And once you're not doing REST, I don't personally think putting HTTP verbs in the method calls doesn't really make it any clearer, and is just likely to confuse matters.
I hear you (REST and batch operations don't mix too well). But wouldn't it be possible to simply provide higher level operations, grouping several of these commands?
> And once you're not doing REST, I don't personally think putting HTTP verbs in the method calls doesn't really make it any clearer, and is just likely to confuse matters.
As I said, it wouldn't necessarily be restricted to HTTP verbs (eg, using 'create' and 'update' instead of 'POST' and 'PUT'), but I do like the idea of separating the resource you act on from the action on this resource.
Do you have an example of the type of operation that would map poorly to REST?
Also interested in hearing if you know anything about GMail's json api? Surely they must have something similar to JMAP which their web client is using even if it the api is proprietary?
However! as someone who has had to debug some particularly gnarly email issues in the past I really do quite like the ability to telnet into an IMAP server to dig around. I also have an appreciation of some of the more horrific things that can go wrong...
So my point is that I'm not sure I agree with "generally much easier to work with than IMAP" - there's two types of "working with". One is building a new client/server from scratch - I'd agree this might be easier, but not something that is done that often. The other is being able to quickly investigate issues with hand-crafted queries, and unfortunately I'd say a simple plain text question-answer interface is better - and here IMAP wins.
Maybe I'm just getting old - it's probably true that almost everyone is happier with HTTP than "telnet-able" protocols (SMTP, IMAP, Redis, etc.) nowadays.
Maybe. I'm curious what the effect would be if a single mobile platform provider (iOS, Android) supported a new protocol with a single main free mail provider. Specifically, Gmail and Android could possibly push major change through.
Look at SPDY, it was originally an experiment with Chrome and only a subset of Google services, and is now being deployed to other properties and iterated on rapidly in other browsers, too.
Honestly, if it's not significantly nicer to work with than IMAP, it won't see traction, and that's fine.
Mind you - MOVE took bloody long enough in IMAP, I was arguing for it for years before it finally got support behind it. There's still not much support for SUBMIT via IMAP though, which means multiple connections via different protocols to do basic emailing, with all the support fun that entails.
Anyway - worst case, we still have a better protocol for our own stuff, and that's not nothing.
Er…
> 5.5. Multiple Commands in Progress
> The client MAY send another command without waiting for the completion result response of a command, subject to ambiguity rules (see below) and flow control constraints on the underlying data stream. Similarly, a server MAY begin processing another command before processing the current command to completion, subject to ambiguity rules.
IMAP already supports batching.
This is a text/event-stream resource, as described in [http://www.w3.org/TR/eventsource/](). The following events are pushed:
- progress: Sent during a long-running api method call to let the client know not to timeout the connection. The data is a JSON object with a single property: `connectionId`, with the id the client sent for that connection. This event does not set a new id in the event source stream.
- push: Sent whenever the user's highest modseq changes. The event id is the new highest modseq. The data for the event is a JSON object, with the following properties:
* **clientId**: `String` (optional). If the change was due to an action initiated by
the API, the `X-ME-ClientId` header will be echoed back as this property. This
allows clients to ignore push events for changes they themselves have made.
* **mailModSeq**: `Number` (The highest modseq for a mailbox)
* **contactsModSeq**: `Number` (The highest modseq for contacts/contact groups)
* **calendarModSeq**: `Number` (The highest modseq for calendar events).The protocol could just as easily go over a WebSocket as HTTP (you'd have to change authentication mechanism, but that would be trivial). The only reason we don't do that at the moment is: 1. Browser support (we support back to IE8, unfortunately) 2. Lack of GZIP compression (particularly of responses from the server); this was being worked on last time I looked, but I haven't checked recently to see if it's been implemented yet. This is crucial for saving bandwidth.
That would get rid of your HTTP overhead; I will put some mention of that in the spec (I think HTTP support should be required, and WebSocket support optional).
As for JSON vs. binary: * There are libraries to use JSON everywhere – binary protocols would require everyone to implement another parser, adding extra difficulty for adoption and a surefire source of bugs. * JSON is readable over the wire, binary is not. Trust me, in the real world, this is a huge advantage for debugging all sorts of things. * JSON can be used trivially in a web app. Apparently this World Wide Web thing might catch on. * The parsing overhead of JSON is not too large; yes it's longer than a binary protocol, but it hasn't been a bottleneck for us, even on mobile.
Parsing latency isn't really an issue; all this stuff is IO bound anyway. CPU cycles are cheap.
A sure-fire way to decrease latency is to send as little as possible.
An empty default browser HTTP request is likely to cost you around 500 bytes in headers, before you have added any headers or data. Contrast that with a binary command which can be 1 byte. And then multiply this by hundreds of requests. CPU cycles on mobile are not cheap. It makes no sense to parse 500 bytes (and then GC this later) when you don't have to.
For instance, if each bucket contains 8 seconds worth of arrival time, then 2^17 buckets mean they'll wrap around only every ~12 days. So if you are syncing after a day away there will be only 8% of the buckets potentially dirty (even if you get an incredible flood of mail!)
Having the bucket use based on time also allows you to avoid some of the roundtrip costs when doing frequent syncing. In a single request, the client can speculatively send its hashes of buckets that are likely to have changed since its last sync time along with its root hash. The server finds which buckets changed, sending those new bucket contents -- it can then see that the top level hash must now agree on both sides and you're done in one RT.
Of course it's possible that other buckets have changed in rare edge cases; in this case the top hashes will not match and you have to do the full Merkle descent to resync.
There are other optimizations you can bring in when you sync the tree:
1. If you are working your way down the remote's tree and you notice that your local signature is zero for the equivalent remote's node signature, then you know that your entire subtree is empty, and you can short-circuit and start downloading entire buckets.
2. Conversely, if you are working your way down the remote's tree and you notice that your local signature is present but the equivalent remote's node signature is zero, then you know that your entire subtree has data, but the remote's entire subtree is empty, and you can short-circuit and start uploading entire buckets.
3. As you work your way down sections of the tree, you can start to build an idea of the average email to bucket ratio. If several buckets are only likely to contain at most one email, then you can short-circuit again.
A fixed-size Merkle tree is a great idea if we wanted to sync the entire set of messages between distributed MX destinations. In fact, we should probably consider that approach for our server <-> server replication. However, JMAP is aimed at client <-> server sync. There are a number of different issues to consider here. The client may have limited or no cache (a webmail app for instance must normally start anew each time, and needs to efficiently fetch just what it needs; if we had to wait for a 60GB mailbox to sync every time we loaded our webmail, well, obviously that's not going to work. A mobile mail client often only wants to download the first few items in the mailbox given the current sort order so it can display them to the client. The JMAP spec allows a client to download, say, just the list of the first 50 messages, newest first, in the Inbox (single round trip). Then, at a later point, it can ask for and receive a space-efficient delta update of what has changed within those first 50 messages in the Inbox, regardless of how big the Inbox really is, or what other messages may have been delivered elsewhere (including lower down the Inbox, but outside the top 50). Again in a single round trip. This is really powerful: you can see that by comparing the refresh time of a folder in the FastMail mobile web app with the iOS or Android native mail app (over IMAP); the mobile web app is much quicker.
The other thing to remember is that the metadata about the message is mutable, and needs to be kept in sync as well. Again, the modseq system described in JMAP allows this to happen efficiently. If you have a distributed server system with multiple masters, you will again use a different system to keep those in sync (as they each assign new modseqs, whereas the client does not). However, this too can be dealt with relatively easily: if the modseq for a message differs between two masters, it must be set on both to max(the higher of the two modseqs, the next highest global modseq on the server with the currently lower modseq). This will ensure that both will end up in the same state, and that the client will get refresh its state for that message if it has to, no matter which master it last synced with.
So essentially, there's a difference in what you want in a protocol for server <-> server synchronisation for distributed mail backends compared to client <-> server sync for mail apps. JMAP is focussed on the latter, but tries to make sure it doesn't include assumptions that severely limit the former. The spec does not require a particular format of message ids: they do not have to be based around IMAP's ascending UIDs (although the sync algorithm for partial mailbox listings may be slightly less efficient if not).
First, the Merkle tree would just be of the IDs not the message contents. I.e. the problem it is trying to solve is for the heavy-weight client that wants to resync its idea of which message UIDs are contained in a mailbox. Which ones it then requests metadata (or full-contents) is up to it.
It's true that for a ephemeral client (like a webapp) you don't even want to get the whole list of 100K messages in your inbox (although you probably will ask the server for a count of them) However, here having a sequential message number hardly helps either. For example, when I click on "sort by date" in my mail client it shows me the messages by sending date, not by what integer IMAP UID they were given.
The problem of "give me the metadata for the top 100 messages sorted by criteria X" is just something that has to be offered by the server. This is no different than features like searching -- if bandwidth were infinite the client could do it, but it's not so you need to do the work near where the data is stored.
The server will have to maintain data structures to make these kinds of queries efficient. The important point however, is that these are conceptually just caches of information that is in the mailstore. Therefore there is no need to keep that synced between far-flung servers.
> The other thing to remember is that the metadata about the message is mutable, and needs to be kept in sync as well
Yes, message data and metadata are separate problems. If two servers both try to update the metadata for message X at the same instant one needs to win (which one is probably not that important) If two servers both try to add new messages to a folder at the same instant they both need to win, ideally without having to coordinate at all in advance.
For the metadata you can use a modseq, a Lamport timestamp, or probably even just a {wallclock time,server ID} tuple. Assuming good time synchronization is reasonable for a mail cluster. If two clients both try to change message flags within 10ms of each other it probably isn't important who wins as long as somebody definitively does.
You suffer from a disease, and it's far too bloody common amongst people who design IETF standards.
You're optimising a rare case, where in the worst case you would be re-iding the bulk of the messages delivered in a time period of a few hours, leading clients to have to re-fetch the metadata again (note: it's only the modseq that would change in our design - the UUID is still globally unique and unchaged).
That's no worse than "FETCH 1:* FLAGS" now.
And in exchange for this slight benefit, you would cause EVERY SINGLE CLIENT EVERYWHERE to implement a significantly more complex protocol which requires 4 roundtrips to resync.
Talk about a dog with a very big, very waggy tail.
I'm sorry, but I can't take this suggestion seriously. I've seen too many protocols which are designed with this sort of focus on the worst cases. You have to make sure they don't totally suck - but the important place to optimise is the common case.
...
It turns out there is a way you can gain almost all of the benefits you're looking for - which is delayed MODSEQ allocation. Newly delivered messages can be created with no MODSEQ. When you sync with another either client or server, you allocate a MODSEQ higher than than any which the other end has seen.
Which means you only pay an extra cost if a JMAP peer or client connects at both ends during the link downtime.