Felix GV created AVRO-2307:
--
Summary: Opt-in setting to improve GC behavior during
deserialization?
Key: AVRO-2307
URL: https://issues.apache.org/jira/browse/AVRO-2307
Project: Apache Avro
Issue Type: Bug
Affects Versions: 1.7.7
Reporter: Felix GV
We have a performance-sensitive project that leverages Avro for an online data
store, and I wanted to report on a couple of Avro deserialization optimizations
we have implemented. On one hand, it is great that Avro’s code is clean and
modular enough to have allowed us to achieve this easily. But on the other
hand, we are leveraging parts of the API which are probably not typically used
by most users, and thus we are exposing ourselves to ongoing maintenance costs
as those “ambiguously-public” APIs might change in future versions. For this
reason, I wanted to gauge the appetite of the Avro community for taking in
those optimizations upstream into the main project.
The minor challenge is that the optimizations we’ve made are not completely
invisible, and therefore should probably be presented as opt-in settings,
rather than new defaults. Below is a summary of both changes.
1. Re-use of byte arrays when instantiating ByteBuffers
When deserializing a byte array that contains a ByteBuffer field, the relevant
portion of the input byte array is copied into a new byte array, which is then
used as the backing array of a new a ButeBuffer.
In our case, we have a few schemas which contain some general metadata and an
opaque byte array payload, which often ends up being a significant portion of
the total byte length. Recopying these bytes results in up to 2x the byte
allocation. The ButeBuffer API, however, provides an alternative behavior where
the backing array can be larger than needed, with an offset and length provided
to indicate the internal boundaries of the payload. In our implementation, we
re-use the input byte array as the ButeBuffer’s backing array, therefore
avoiding a copy.
The caveat in this case is that this only works properly for use cases that
don’t mutate the content of the bytes (neither the input nor the deserialized
object). In our case this assumption is valid.
If this was implemented in the open-source project, there are a few ways this
could be achieved:
# There could be a config flag on the decoder or elsewhere that allows a user
to opt-in to this mode. In this case, it may be safer to return a special
read-only ButeBuffer implementation that throws an exception if any mutation is
attempted, indicating that the flag ought to be turned off to support mutations.
# It could be the default mode, but wrapped in a modified ByteBuffer
implementation which defers the copy of the content lazily until (and only if)
one of the mutation API is called.
Either way, this requires a custom ByteBuffer implementation with special
behavior in order to be fully clean and safe, however, in the first approach,
the default behavior would still return regular ByteBuffer instances.
2. Primitive (non-boxing) implementation of lists
Another challenge we’ve come across is that lists of primitive types (floats in
our case) are always boxed into Object Floats by Avro. In our case, this
resulted in millions of Objects / second getting allocated, causing
pathological GC pressure.
To solve this, we have implemented an alternative version of the Avro Array
class, but which instead of hanging on to an array of generically-typed
Objects, internally, hangs on to an array of primitive floats. This causes no
boxing at deserialization time, but there is a further challenge which is that
since Avro array fields are exposed as Java Lists, the regular functions of the
API all return Objects, therefore merely deferring the boxing to a slightly
later point in time. To get around this further complication, we have added a
getPrimitive(i) function which returns primitive items directly. In order to be
able to use this more optimized function, it is necessary for us to cast the
list into our own type, otherwise we wouldn’t see the new function. The end
result is quite dramatic, performance-wise, reducing our p99 latencies down to
a quarter to a third of their original values.
One challenge here is that the “PrimitiveFloatArray” class is an almost
complete copy of the Array class, basically just stripping away the generics.
If we were to contribute this upstream to the open-source project, I imagine we
might want to do this not only for floats but for boolean, int, long and double
arrays as well. This would mean roughly 5x the same copy-pasted implementation,
which is not ideal from a maintenance standpoint. The generic types are nicer
in that sense, but unfortunately, Java generics do not support primitives. In
our case, we are willing to pay that maintenance cost in exchange for the
dramatic GC reduction it gives