[infinispan-dev] Stream encoding of Flags and future compatibility

Dan Berindei dan.berindei at gmail.com
Tue May 15 15:39:20 EDT 2012


On Tue, May 15, 2012 at 1:11 PM, Sanne Grinovero <sanne at infinispan.org> wrote:
> On 15 May 2012 09:36, Dan Berindei <dan.berindei at gmail.com> wrote:
>> On Mon, May 14, 2012 at 2:11 PM, Sanne Grinovero <sanne at infinispan.org> wrote:
>>> On 11 May 2012 22:30, Dan Berindei <dan.berindei at gmail.com> wrote:
>>>> On Fri, May 11, 2012 at 7:23 PM, Sanne Grinovero <sanne at infinispan.org> wrote:
>>>>> On 11 May 2012 16:37, Galder Zamarreño <galder at redhat.com> wrote:
>>>>>> Quickly tried this and caused no issues:
>>>>>> https://github.com/galderz/infinispan/commit/7718926e5a4a6763506250362d7bd5cbdccd2931
>>>>>
>>>>> Looks good! I'm sure this doesn't solve all future migration problems,
>>>>> but if we could keep this kind of tricks around it should improve
>>>>> odds.
>>>>> IMHO, this is a kind of sensitivity that we should apply across all
>>>>> areas (not just flags).
>>>>>
>>>>
>>>> Looks interesting, but then you have the opposite problem: not all new
>>>> flags can be ignored, so you need a way to specify that a new flag is
>>>> "required". E.g. if we had just added a ZERO_LOCK_ACQUISITION_TIMEOUT
>>>> flag then the client would be expecting spurious failures, but not
>>>> extra long delays.
>>>
>>> You're right - but did you read the conversation on github? We already
>>> pointed this out, still I believe we should have an option to ignore
>>> unknown flags if/when/exclusively we think the migration is safe: we
>>> should be able to tell after the fact, possibly even write migration
>>> tests, but can't predict the future.
>>>
>>> We could also use a single bit in the externalized representation of a
>>> flag to mean "safe to be ignored" for any flag, but I'm not sure that
>>> all cases would be black/white .. more likely it will depend on use
>>> case or actual configuration.
>>>
>>
>> Right, the flag serialized format should tell the receiver whether it
>> can be ignored. But if we encode an EnumSet<Flag> as a bitset, each
>> flag will have only one bit, so we wouldn't have a place for the
>> "migration safe" bit. I guess we could send two bit sets, one with the
>> flags and one with the "migration safe" bits...
>
> I know I proposed this myself, but I would rather avoid this path..
> even if some flags are unsafe, they might be safe in the configuration
> / use case, and we might want to force it to proceed nevertheless. I
> think we should always have the option to "force it to proceed".
> In worst case one might want to block all transactions and kill the
> clients, but you have to propose a way to migrate the data.
>

I'm starting to warm up to the idea of making all flags "migration
safe" myself... in the end, it's the user's decision, we just have to
make sure that he knows what to expect.

> #On two bitsets being sent..
>
> First problem is that we're sending redundant information; second is
> that it should not be responsibility of the older node (in terms of
> version of the code being run) to properly understand the newer nodes,
> it should be the other way around.
>
> So rather than sending two bitsets, in this case the responsibility
> belongs to the *sender*: since it's the one containing the newer code,
> it might know how to deal with older nodes, or might be patched/fixed
> to know how to deal with older nodes. It's obviously harder to patch
> existing nodes! Especially if they are running already (and you don't
> have Byteman).
> If we could keep track of the version of other nodes in the View, we
> could have a translation table plugged in on how to talk to a
> different version node.. maintaining such a table would be much
> simpler as it would live independently from the main code, and be
> fixed and hacked on.. possibly even loaded dynamically. So that's what
> I would do in the long term.. in the short term, let's just save some
> bytes using optimal encoding for the EnumSet.
>

Keeping track of the versions on each node and sending different
things is certainly doable, but I fear in replicated mode it would
force us to replace broadcasts with multiple unicasts. So there is a
small risk that it will bring the cluster down anyway due to extra
traffic ;)

The extra traffic during the upgrade is probably acceptable, I'm
actually more concerned about the extra checks that we'd have to do
all the time to ensure that we are not doing an upgrade at the moment
and we don't need the special treatment...

> Especially of note, if the sender (newer node) knows by looking into
> the translaction table that an operation/command can't be possibly
> sent to the older node, it should be easier to find an alternative or
> workaround.
> when migrating an existing cluster by a rolling upgrade, for sure the
> existing application is NOT using features which exist only in the
> never version.. that's granted and we can take advantage of that.
>
>>
>>>
>>>>> On a totally different page, why are we serializing Flags one-by-one ?
>>>>> We mostly need to serialize EnumSets right?
>>>>> An EnumSet can be encoded by using the bits of a couple of bytes.
>>>>> Three bytes looks like enough for all our needs.. we could even be
>>>>> clever and reserve a special Externalizer-ID for the empty set, to
>>>>> avoid 3 bytes where none are needed.
>>>>> While currently we need an integer (4 bytes) to encode the header for
>>>>> "EnumSet", plus (4 bytes header + 1 byte value) * each flag -> a lot.
>>>>>
>>>>
>>>> RiverMarshaller already has an optimization for the empty set:
>>>> https://github.com/dmlloyd/jboss-marshalling/blob/master/river/src/main/java/org/jboss/marshalling/river/RiverMarshaller.java#L613
>>>
>>> That code makes perfectly sense in a general purpose use case, but it
>>> still needs to serialize the Class definition: we can avoid that, so
>>> we should!
>>>
>>
>> I think we need to encode the Flag class name as long as we consider
>> the flag set as just another parameter in the Object[] parameter array
>> - so the externalizer doesn't know how to distinguish between an
>> EnumSet<Flag> and an EnumSet<UserEnumType>.
>
> Can't we write it using the special-purpose Externalizer explicitly?
> To read it, it will have it's own Externalizer ID.
>

When writing, the marshaller can only use the object's class to find
the appropriate externalizer for an object, and I'm pretty sure
getClass() will return the same thing for EnumSet<Flag> and
EnumSet<UserEnumType>.

>
>>>> I'm not sure why it doesn't encode each element as a bit, it might be
>>>> to keep wire compatibility when the order of values in an enum
>>>> changes.
>>>
>>> That's a safe behaviour, expected for a default use case. But if we
>>> decide to add the UNKNOWN flag, we could use bitsets.
>>>
>>
>> I think the UNKNOWN flag is useful as long as we parse each flag
>> independently, but if we start encoding the flags as a bitset (or
>> two), then we don't need UNKNOWN any more. We would only need to
>> ensure the position of a flag in the bitset remains constant between
>> versions, which means either relying on the declaration order and
>> keeping that constant (no removing or reordering of flags) or adding a
>> position field to the enum.
>
> Right, I only asked for the UNKNOWN for the time being, as we don't
> have anything better yet.
> And I guess it's safe to have such a flag for future.. you never know.
>
>> Assuming the order in the bitset remains constant from one version to
>> the next, there are only two scenarios:
>> a) The added flags are not important, and the receiver can just ignore
>> them - in which case the parsed flag set will only include the flags
>> that the receiver knows about.
>> b) The added flags are important, and the receiver should reject the
>> command - in which case the externalizer should just throw an
>> exception instead of reading the flag set and setting the UNKNOWN
>> flag.
>>
>>
>>>> However, because there is only one EnumSet for all Enum types, a
>>>> hypothetical EnumSetExternalizer also needs to write the name of the
>>>> enum class - if we wanted to serialize EnumSet<Flag> in 2 bytes then
>>>> we'd need to make the transformation in ReplicableCommandExternalizer.
>>>>
>>>> Cheers
>>>> Dan
>>>>



More information about the infinispan-dev mailing list