[
https://issues.apache.org/jira/browse/AVRO-2164?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16523603#comment-16523603
]
Andy Coates commented on AVRO-2164:
-----------------------------------
I guess the main issue that I see is that the current decimal implementations
in Avro doesn't allow for schema evolution where the scale changes. Consider a
use-case where a schema has a decimal field with scale 4. Then at some point
later it comes to light that 4 is too little, and a scale of 5 is needed, (or
it's too much and is wasting space and only 3 is needed). With the current
implementation, if the scale is changed, then reading the a record serialised
with the old schema using the new schema as a read-schema will result in the
wrong value. This, IMHO, is a big issue. This is data corruption.
Possible ways to solve this:
- don't allow scale changes, i.e. read/write schemas with different scale
should be considered incompatible.
- covert values on the fly, i.e. deserialize the decimal using the old scale
and attempt to set the new scale. Where the new scale is larger this will
always work. Where the new scale is smaller this may through an exception. (It
may also make sense to allow the user to define rounding behaviour in the case
where the scale has reduced).
- encode the scale in the serialised form, i.e. create a new type, (either
first class or logical), where the serialised form is prefixed with the scale.
Of these, maybe the best is to convert on the fly?
> Make Decimal a first class type.
> --------------------------------
>
> Key: AVRO-2164
> URL: https://issues.apache.org/jira/browse/AVRO-2164
> Project: Avro
> Issue Type: Improvement
> Components: logical types
> Affects Versions: 1.8.2
> Reporter: Andy Coates
> Priority: Major
>
> I'd be interested to hear the communities thoughts on making decimal a first
> class type.
> The current logical type encodes a decimal into a _bytes_ or _fixed_. This
> encoding does not include any information about the scale, i.e. this encoding
> is lossy.
> There are open issues around the compatibility / evolvability of schemas
> containing decimal logical types, (e.g. AVRO-2078 & AVRO-1721), that mean
> reading data that was previously written with a different scale will result
> in data corruption.
> If these issues were fixed, with suitable compatibility checks put in place,
> this would then make it impossible to evolve an Avro schema where the scale
> needs to be changed. This inability to evolve the scale is very restrictive,
> and can result in high overhead for organizations that _need_ to change the
> scale, i.e. they may potentially need to copy their entire data set,
> deserializing with the old scale and re-serializing with the new.
> If _decimal_ were promoted to a first class type, this would allow the scale
> to be captured in the serialized form, allow for schema evolution support.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
