Is there a way to do batching in that transformation? I'm assuming for now no. or may be using in conjuntion with GoupIntoBatches
On Mon, Apr 15, 2024 at 9:29 AM Ruben Vargas <[email protected]> wrote: > > Interesting > > I think the cache feature could be interesting for some use cases I have. > > On Mon, Apr 15, 2024 at 9:18 AM XQ Hu <[email protected]> wrote: > > > > For the new web API IO, the page lists these features: > > > > developers provide minimal code that invokes Web API endpoint > > delegate to the transform to handle request retries and exponential backoff > > optional caching of request and response associations > > optional metrics > > > > > > On Mon, Apr 15, 2024 at 10:38 AM Ruben Vargas <[email protected]> > > wrote: > >> > >> That one looks interesting > >> > >> What is not clear to me is what are the advantages of using it? Is > >> only the error/retry handling? anything in terms of performance? > >> > >> My PCollection is unbounded but I was thinking of sending my messages > >> in batches to the external API in order to gain some performance > >> (don't expect to send 1 http request per message). > >> > >> Thank you very much for all your responses! > >> > >> > >> On Sun, Apr 14, 2024 at 8:28 AM XQ Hu via user <[email protected]> > >> wrote: > >> > > >> > To enrich your data, have you checked > >> > https://cloud.google.com/dataflow/docs/guides/enrichment? > >> > > >> > This transform is built on top of > >> > https://beam.apache.org/documentation/io/built-in/webapis/ > >> > > >> > On Fri, Apr 12, 2024 at 4:38 PM Ruben Vargas <[email protected]> > >> > wrote: > >> >> > >> >> On Fri, Apr 12, 2024 at 2:17 PM Jaehyeon Kim <[email protected]> wrote: > >> >> > > >> >> > Here is an example from a book that I'm reading now and it may be > >> >> > applicable. > >> >> > > >> >> > JAVA - (id.hashCode() & Integer.MAX_VALUE) % 100 > >> >> > PYTHON - ord(id[0]) % 100 > >> >> > >> >> Maybe this is what I'm looking for. I'll give it a try. Thanks! > >> >> > >> >> > > >> >> > On Sat, 13 Apr 2024 at 06:12, George Dekermenjian <[email protected]> > >> >> > wrote: > >> >> >> > >> >> >> How about just keeping track of a buffer and flush the buffer after > >> >> >> 100 messages and if there is a buffer on finish_bundle as well? > >> >> >> > >> >> >> > >> >> > >> >> If this is in memory, It could lead to potential loss of data. That is > >> >> why the state is used or at least that is my understanding. but maybe > >> >> there is a way to do this in the state? > >> >> > >> >> > >> >> >> On Fri, Apr 12, 2024 at 21.23 Ruben Vargas <[email protected]> > >> >> >> wrote: > >> >> >>> > >> >> >>> Hello guys > >> >> >>> > >> >> >>> Maybe this question was already answered, but I cannot find it and > >> >> >>> want some more input on this topic. > >> >> >>> > >> >> >>> I have some messages that don't have any particular key candidate, > >> >> >>> except the ID, but I don't want to use it because the idea is to > >> >> >>> group multiple IDs in the same batch. > >> >> >>> > >> >> >>> This is my use case: > >> >> >>> > >> >> >>> I have an endpoint where I'm gonna send the message ID, this > >> >> >>> endpoint > >> >> >>> is gonna return me certain information which I will use to enrich my > >> >> >>> message. In order to avoid fetching the endpoint per message I want > >> >> >>> to > >> >> >>> batch it in 100 and send the 100 IDs in one request ( the endpoint > >> >> >>> supports it) . I was thinking on using GroupIntoBatches. > >> >> >>> > >> >> >>> - If I choose the ID as the key, my understanding is that it won't > >> >> >>> work in the way I want (because it will form batches of the same > >> >> >>> ID). > >> >> >>> - Use a constant will be a problem for parallelism, is that correct? > >> >> >>> > >> >> >>> Then my question is, what should I use as a key? Maybe something > >> >> >>> regarding the timestamp? so I can have groups of messages that > >> >> >>> arrive > >> >> >>> at a certain second? > >> >> >>> > >> >> >>> Any suggestions would be appreciated > >> >> >>> > >> >> >>> Thanks.
