Retry thrift calls when decoding fails due to retryable error

[prashantv]

Apache Thrift wraps errors as a string, which makes it hard to get to the underlying error.

We see errors like:

*[struct] error reading struct: *[struct] error reading struct: *[struct] error reading struct: error reading field 4: tchannel error ErrCodeNetwork: tchannel: socket closed, remote closed

The CanRetry method does not unwrap these errors (since there's no strong type currently), which means that a network error that happens during thrift decoding will not be retried.

This only tends to affect large calls (>64kb) since calls less than 64kb fit in a frame, and thrift decoding only starts after processing the headers in the first frame.

[hanjunx]

lgtm Hanjun

…

On Wed, Feb 1, 2017 at 7:05 AM, Prashant Varanasi ***@***.***> wrote: Apache Thrift wraps errors as a string, which makes it hard to get to the underlying error. We see errors like: *[struct] error reading struct: *[struct] error reading struct: *[struct] error reading struct: error reading field 4: tchannel error ErrCodeNetwork: tchannel: socket closed, remote closed The CanRetry method does not unwrap these errors (since there's no strong type currently), which means that a network error that happens during thrift decoding will not be retried. This only tends to affect large calls (>64kb) since calls less than 64kb fit in a frame, and thrift decoding only starts after processing the headers in the first frame. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#575>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACW3y1wSAFWDhnFw1HsidG7lK2xJTV0Fks5rX75IgaJpZM4LzSjE> .