thrift-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mayan Moudgill <ma...@bestweb.net>
Subject Re: heterogeneous collections
Date Tue, 04 May 2010 18:27:48 GMT
Hmmm...
1. The debate is about sending ONLY field sizes OR sending type 
identifiers for lot of objects.

  Consider the encoding of
struct {1:i32 A, 2: list<i16> {B0,B1,B2}
In TBinaryProtocol it gets encoded as:
   T_STRUCT 1 T_I32 a_val 2 T_LIST 3 T_I16 b0_val b1_val b2_val T_STOP

With a strongly typed system where the only thing allowed is different 
numbers of fields, the encoding would become
   1 B=4 a_val 2 B=6 3 b0_val b1_val b2_val T_STOP
where B= are the byte counds for the fields.

2. In thrift, we need the byte counts to skip unknown fields. So, we 
don't really need to know the actual byte count, but we need to know 
when a field will end. Note also that during the demarshalling process, 
at any one time, there will be at most one field being skipped. Also, 
fields can be managed as a stack.

Assume one wanted to assemble data in chunks of 64KB. In that case, when 
a block 64KB had been assembled, some fields might still be being 
written. Of these fields, some would get fully written ("completed") in 
the next 64KB, and some would still be pending. We will keep information 
about the fields completed in a 64KB block in the beginning of the 
block. The completed-field information field will look like:
* Depth of stack at beginning of block
* number of pops
* list of byte-offset

So:
- writeFieldBegin() pushes field onto a stack, and reserves 2B for the 
size; intially set value to 65535
- writeFieldEnd() pops field from a stack; if the entire field is 
contained in a 64KB block, write byte offset  to field. If the field is 
from previous block, append byte offset to completed-information field.

When skipping, either the skip is > 65535, in which case skip to offset 
within block, or it is 65535, in which case skip to the next block, read 
the completed-field info, see if the skipped field would get popped off 
(i.e. completed). If yes, skip to the offset recorded in the byte-offset 
list, else skip to end of block and repeat.


Mark Slee wrote:
> Right, but this requires you to store state in the protocol (currently,
> protocols *can* do this but it is not a requirement), and also mandates
> the use of an internal buffer *at least* as large as the largest complex
> object you are going to serialize (there is no way to tell a socket API
> to go back and fill in 4 bytes earlier in the sequence... that means the
> whole serialized object has to go in application-space memory before
> writing to the transport). Not ideal if you plan to serialize very large
> containers. This was a consciously considered issue that we wanted to
> avoid.
> 
> If you really do "prefer 2D array of arrays," then you truly are a C
> programmer at heart, and we may just have to chalk this up as having
> different opinions about simplicity. =)
> 
> Honestly, though, this debate seems silly to me. Whether it's a
> field-size or a type-identifier, it's still an extra byte of metadata
> (or more if the length doesn't fit in a byte) that the protocol needs to
> encode in addition to field-identifiers. The only difference is that one
> describes the application content (field-type), the other describes the
> serialized data.


> Cheers,
> mcslee
> 
> -----Original Message-----
> From: Mayan Moudgill [mailto:mayan@bestweb.net] 
> Sent: Monday, May 03, 2010 11:00 PM
> To: thrift-user@incubator.apache.org; Mark Slee
> Subject: Re: heterogeneous collections
> 
> 
> I wouldn't have worried too much about the encoding; you can do it using 
> pretty much the same write*() interface with the following caveats:
> - all writes are to an "array" (generally, I prefer a 2D array of arrays)
> - writeFieldBegin pushes the offset within the array and advances by 4 bytes
> - write FieldEnd pops the offset, subtracts it from current offset and 
> writes that value (=field bytes) to the offset.
> 
> I think you overstate the "simplicity" of the current scheme.
> 
> 
> Mark Slee wrote:
> 
> 
>>The protocol scheme was written the way it was because it was very
>>simple, transparent, straightforward to implement, safe to version
>>changes, and reasonably defensive.
>>
>>- Field identifiers are necessary for versioning Type identifiers
>>- are necessary so that we know how to skip fields that we don't
>>- recognize Therefore, the protocol sends a field identifier, then a
>>- type identifier, then the data
>>
>>We could have used field-size instead of a type identifier. That
>>simplifies the skip read-operation, but comes at the cost of making
>>the write operation much more complicated. It means that if you are
>>serializing a complex type, you have to first compact the whole thing
>>down to determine its total size in bytes, then write out that header.
>>This leads to internal-buffering code in the protocols, not fun when
>>you're dealing with containers-of-containers. Even simple cases are
>>awkward, I can't know the byte-length of a list of strings without
>>actually iterating over all the strings first.
>>
>>So, using the type-identifier system keeps the TProtocol interface
>>incredibly flat and obvious, just serial calls to read/write simple
>>values, always one at a time.
>>
>>The checking of types for known field ids is just basic defensiveness,
>>protection against someone changing the type of a field but forgetting
>>to update its identifier. We don't generate errors because this is
>>considered the same class of occurrence as an invalid field identifier.
>>
>>I totally agree with you about Thrift *seeming* like a partial attempt
>>to implement dynamic RPC. This was basically my point -- I know they
>>look similar -- so I do happily excuse people for thinking this. =)
> 
> 
>>Cheers,
>>mcslee
>>
>>-----Original Message-----
>>From: Mayan Moudgill [mailto:mayan@bestweb.net] 
>>Sent: Monday, May 03, 2010 9:38 PM
>>To: thrift-user@incubator.apache.org; Mark Slee
>>Subject: Re: heterogeneous collections
>>
>>
>>If the goal of Thrift was to transport strongly typed data, then it begs 
>>the question: why was the protocol scheme being used currently adopted?
>>
>>Clearly, if the data is typed with the types being agreed to at both 
>>ends, then NO type information needs to be exchanged (other than you 
>>container lengths - which you may not consider type information). If the 
>>data is typed, but the there may be disagreement over the type at the 
>>reciever, then you have to send the complete type information along with 
>>the data. TBinaryProtocol does neither.
>>
>>Thrift does't have strict strict typing: the stated goal of thrift to 
>>support versioning, where the transmitted type and the receiver type are 
>>permitted to differ by the addition or deletion of fields. This means 
>>that the only information that needs to be transmitted are field-id and 
>>field-size; either the field is a known one, in which case full type 
>>info is known, or the field is unknown, in which case the number of 
>>bytes to skip is known.
>>
>>Assuming that Thrift were intended to be strongly typed, the only reason 
>>to actually transmit as much type information as TBinaryProtocol 
>>actually does is that implicitly Thrift is also allowing for the type of 
>>fields to be changed. Was this intended to account for the case where a 
>>field was deleted, then reused [which does beg the question - what 
>>happens if the reused type is the same as the original?]; if so, there 
>>may very well have been different and better ways of doing this.
>>
>>Other than that the only reason I can come up with is that this was some 
>>kind of type-checking half-measure to ensure correctness. But the 
>>default behavior on a type-mismatch appears to be to discard the field, 
>>not generate some kind of error.
>>
>>So, given the implementation of TBinaryProtocol, people could be excused 
>>for thinking that its a partial attempt to implement a dynamically typed 
>>RPC.
>>
>>Mark Slee wrote:
>>
>>
>>>>>If, however, you're encoding the data for demarshalling at the server,

>>>>>it sounds like you want a different RPC framework.
>>>
>>>I'm going to slightly hijack the conversation to wax philosophic for a
>>>minute here. I think this statement roughly captures my sentiment here.
>>>
>>>One of Thrift's early goals was basically to do just one thing, but do
>>>it very simply and efficiently across lots of platforms. That thing is
>>>*strongly-typed* RPC and data-serialization. All of the components were
>>>essentially designed under the assumption that they would always be
>>>strongly-typed, and that they should always map to something efficient
>>>and obvious in a language like C++.
>>>
>>>Now, a lot of the things Thrift does are very *similar* to other
>>>sorts of interesting mechanisms data-serialization, marshalling,
>>>containering, and whatnot. I think it can be very tempting to look at
>>>these similarities, analyze the distance between the two things, and
>>>decide since that distance looks pretty crossable, so we should just
>>>build a bridge to connect the two.
>>>
>>>My fear is that in the long run this turns a small, neat, island into
>>>a complicated mess of bridges. If you find the right viewing angle and
>>>it's not a foggy day, you can sometimes still see the little island
>>>underneath the bridges, but this Thrift thing definitely looks like
>>>bridgework, not an island.
>>>
>>>In the long term, my personal bias is that this is bad for Thrift. Most
>>>people interested in building these features need them to solve specific
>>>problems and only care about one or two target languages. If we do a lot
>>>of this, we end up with a patchwork set of variable feature-lists that
>>>are inconsistent across languages. The Thrift "brand" will invariably
>>>move away from "simple, lightweight, lets you do the same thing in all
>>>programming languages" towards "a bit complicated, does some things in
>>>some languages."
>>>
>>>Part of the idea of Thrift's modular transport/protocol design
>>>was that it would make it easy for people to implement custom
>>>extensions/modifications to the system *outside of the core project.*
>>>Want to sub in your own weird encoding/transport/whatever? No problem,
>>>just write a TProtocol. Think other folks will be into it? Cool, post
>>>it online and send an email to the thrift-user@ list. Turns out lots of
>>>folks want to use it? Then maybe we should incorporate it.
>>>
>>>For better or worse, I really think simple things like "how many source
>>>files appear to be in this tarball?" can matter a lot for software
>>>adoption. Even if a project is just 10 easy-to-read files at its core,
>>>when you have to locate those 10 files amongst 40 files of extensions
>>>and add-ons, and the default make configuration builds everything, the
>>>project starts feeling like a complicated, awkward thing to deal with,
>>>and us engineers start getting that itchy feeling of "I can't possibly
>>>understand this entire thing, surely it is too complicated and slow, why
>>>don't we just write our own from scratch."
>>>
>>>I don't expect everyone to agree with this, and the direction of the
>>>project is ultimately at the behest of the developers most actively
>>>working on it, but when it comes to things like dynamic or heterogeneous
>>>containers, my opinion is that they just shouldn't be a core part of a
>>>strongly-typed software project with stated simplicity goals.
>>
>>
>>
>>>Cheers,
>>>Mark
>>>
>>>-----Original Message-----
>>>From: Mayan Moudgill [mailto:mayan@bestweb.net] 
>>>Sent: Monday, May 03, 2010 10:03 AM
>>>To: thrift-user@incubator.apache.org; alex@bizo.com
>>>Subject: Re: heterogeneous collections
>>>
>>>
>>>The idea of marshalling to strings seems somewhat counter-productive; 
>>>after all, you're marshalling the data using Thrift, which then gets 
>>>sent to a server, and demarshalls it. Now, on top of that you're adding 
>>>another layer of marshalling.
>>>
>>>A similar thing happens in  Cassandra (except that they use binary 
>>>instead of strings), but at least at Cassandra the user-marshalled data 
>>>is uninterpreted at the server - it only handles the data as an 
>>>uninterpreted blob, so the marshalling/demarshalling is only confined to 
>>>the client [I still wonder about how version control is managed - does 
>>>everyone end up rolling their own?]
>>>
>>>If, however, you're encoding the data for demarshalling at the server, 
>>>it sounds like you want a different RPC framework. For instance, do you 
>>>really need the version flexibility that is provided by Thrift? Are your 
>>>types fixed at source & destination? Do you need a leaner transport? In 
>>>fact, why did you pick Thrift in the first place?
>>>
>>>Apropos the discussion on scalar/string compression in the 
>>>https://issues.apache.org/jira/browse/THRIFT-110
>>>I'm curious: if a particular application would tend to compress better 
>>>using a different algo than the one(s) provided, what happens?
>>>
>>>
>>>
>>>
>>>>On Mon, May 3, 2010 at 7:09 AM, Bryan Duxbury <bryan@rapleaf.com> wrote:
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>>There is already a totally viable workaround, though - make a Union of
the
>>>>>types you want in your collection, and then make the field list<YourUnion>.
>>>>>You get basically all the capabilities with very few drawbacks, plus the
>>>>>ability to include multiple logical "types" in the collection, not just
>>>>>physical types. Of course, if you literally need "any" possible object
to
>>>>>go
>>>>>into the collection, then this won't do it for you.
>>>>>
>>>>
>>>>
>>>>Thanks for the suggestion, Bryan.
>>>>
>>>>I'm experimenting with marshalling my values to strings (I only deal with
>>>>basic types such as int32, int64, strings) right now.   If that doesn't
>>>>work, I'll go with your suggestion.
>>>>
>>>>alex
>>>>
>>>
>>>
>>>
>>>
>>
>>
> 
> 
> 


Mime
View raw message