[erlang-questions] "actor database" - architectural strategy question
Mon Feb 17 22:46:33 CET 2014
On Mon, Feb 17, 2014 at 9:22 PM, Miles Fidelman
> Joe Armstrong wrote:
>> This sounds interesting. To start wit, I think swapping processes to
>> disk is just an optimization.
>> In theory you could just keep everything in RAM forever. I guess
>> processes could keep their state in dictionaries (so you could roll them
>> back) or ets tables (if you didn't want to roll them back).
>> You would need some form of crash recovery so processes should write some
>> state information
>> to disk at suitable points in the program.
> Joe... can you offer any insight into the dynamics of Erlang, when
> running with large number of processes that have very long persistence?
No - this area has not to my knowledge been investigated. The "use lots of
processes" or "as many processes as necessary" has an implicit assumption
that a) the processes are not very large and
b) not very long lived. At the back of my mind I'm thinking of a) as "a few
hundred KB resident size" and
b) a few seconds to minutes. I'm *not* thinking MBs and years. The latter
requirements fit into our
"telecoms domain" - a few thousands to tens of thousands of computations
living for "the length of a telephone call" ie (max) hours but not years.
Some kind of "getting things out of memory and onto disk when not needed"
layer is needed for your problem,
> Somehow, it strikes me that 100,000 processes with 1MB of state, each
> running for years at a time, have a different dynamic than 100,000
> processes, each representing a short-lived protocol transaction (say a web
My first comment is, thanks for providing some numbers :-) I keep saying
time and time
again, don't ask questions without numbers. 100K processes with 1MB of
state = 10^11 bytes
so you'd need a really big machine to do this. Assuming say 8GB of memory
and 1MB of state
you'd have an upper limit of 8K processes. This assumes a regular spinning
disk. I guess if you have
a big SSD the story changes.
So you either have to reduce the size of the state, or the number of
processes. The state can (I suppose) be partitioned into a (small) index
and a (larger) content. So I'd keep the index in memory and the content
on disk (or cached).
> Coupled with a communications paradigm for identifying a group of
> processes and sending each of them the same message (e.g., 5000 people have
> a copy of a book, send all 5000 of them a set of errata; or send a message
> asking 'who has updates for section 3.2).
Hopefully all 5000 people will not want the errata at the same time
> In some sense, the conceptual model is:
> 1. I send you an empty notebook.
> 2. The notebook has an address and a bunch of message handling routines
> 3. I can send a page to the notebook, and the notebook inserts the page.
> 4. You can interact with the notebook - read it, annotate it, edit certain
> sections - if you make updates, the notebook can distribute updates to
> other copies - either through a P2P mechanism or a publish-subscribe
> At a basic level, this maps really well onto the Actor formalism - every
> notebook is an actor, with it's own address. Updates, interactions,
> queries, etc. are simply messages.
> Since Erlang is about the only serious implementation of the Actor
> formalism, I'm trying to poke at the edge cases - particularly around
> long-lived actors. And who better to ask than you :-)
It's a very good question. I like questions that poke around at the edges
of what is possible :-)
> In passing: Early versions of Smalltalk were actor-like, encapsulating
> state, methods, and process - but process kind of got dropped along the
> way. By contrast, it strikes me that Erlang focuses on everything being a
> process, and long-term persistence of state has taken a back seat.
Yes - I guess the real solution would be to change the scheduler to swap
processes to disk after they had waited for more than (say) 10 minutes for
a message, and resurrect them when they are sent a message.
The idea that they might be swapped out for years hadn't occurred to me.
> I'm trying to probe the edge cases. (I guess another way of looking at
> this is: to what extent is Erlang workable for writing systems based around
> the mobile agent paradigm?)
Pass - at the moment you'd have to implement you own object layer to do
I guess you could do this yourself by making send and receive library
making the state of a process explicit rather than implicit, then slicking
a large store (like riak). If you cache the active processes in memory this
>> What I think is a more serious problem is getting data into the system in
>> the first place.
>> I have done some experiments with document commenting and annotation
>> systems and
>> found it very difficult to convert things like word documents into a form
>> that looks half
>> decent in a user interface.
> Haven't actually thought a lot about that part of the problem. I'm
> thinking of documents that are more form-like in nature, or at least built
> up from smaller components - so it's not so much going from Word to an
> internal format, as much as starting with XML or JSON (or tuples), building
> up structure, and then adding presentation at the final step. XML -> Word
> is a lot easier than the reverse :-)
> On the other hand, I do have a bunch of applications in mind where parsing
> Word and/or PDF would be very helpful - notably stripping requirements out
> of specifications. (I can't tell you how much of my time I spend manually
> cutting and pasting from specifications into spreadsheets - for
> requirements tracking and such.) Again, presentation isn't that much of an
> issue - structural and semantic analysis is. But, while important, that's
> a separate set of problems - and there are some commercial products that do
> a reasonably good job.
> I want to parse Microsoft word files and PDF etc. and display them in a
>> format that is
>> recognisable and not too abhorrent to the user. I also want to allow
>> on-screen manipulation of
>> documents (in a browser) - all of this seems to require a mess of
>> Before we can manipulate documents we must parse them and turn them into
>> a format
>> that can be manipulated. I think this is more difficult that the storing
>> and manipulating documents
>> problem. You'd also need support for full-text indexing, foreign language
>> and multiple character sets and so
>> on. Just a load of horrible messy small problems, but a significant
>> barrier to importing large amounts
>> of content into the system.
>> You'd also need some quality control of the documents as they enter the
>> system (to avoid rubbish in rubbish out), also to maintain the integrity of
>> the documents.
> Again, for this problem space, it's more about building up complex
> documents from small pieces, than carving up pre-existing documents. More
> like the combination of an IDE and a distributed CVS - where fully
> "compiled" documents are the final output.
>> If you have any ideas of now to get large volumes of data into the system
>> from proprietary formats
>> (like ms word) I'd like to hear about it.
> Me too :-) Though, I go looking for such things every once in a while,
> - there are quite a few PDF to XML parsers, but mostly commercial ones
Suck - then you have to buy them to find out if they are any good
> - there are a few PDF and Word "RFP stripping" products floating around,
> that are smart enough to actually analyze the content of structured
> documents (check out Meridian)
> - later versions of Word export XML, albeit poor XML
> - there are quite a few document analysis packages floating around,
> including ones that start from OCR images - but they generally focus on
> content (lexical analyis) and ignore structure (it's easier to scan a
> document and extract some measure of what it's about - e.g. for indexing
> purposes; it's a lot harder to find something that will extract the outline
> structure of a document)
> In theory, there is no difference between theory and practice.
> In practice, there is. .... Yogi Berra
> erlang-questions mailing list
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the erlang-questions