Comments (8)
maybe seconds will be more flexible for "softer" use cases
from datatrove.
We encounter them often, usually quite a lot of tasks fail (even with randomize start). We simply relaunch them afterwards.
I believe the requests from our cluster all come from the same IP, so I do not really have any recommendations on this
from datatrove.
Super interesting! Were these network speed improvements (to s3 I assume?) only during the beginning of the job (when each worker is listing all the files) or throughout it?
Also, what logs did you have in mind? Maybe you could add them to your upcoming PR?
from datatrove.
I think it's a good idea. Would you be interested in creating a PR for this change? Maybe randomize_start
could become randomize_start_duration
with a default value of 0 (disabled)
from datatrove.
Iād be happy to create a PR for this change. Regarding the unit for randomize_start_duration, do you think it would be more appropriate to use seconds or minutes as the default unit of time?
from datatrove.
Regarding this issue, did you also encounter Too Many Requests
errors while working with Common Crawl?
Iām interested in determining the optimal number of multi-processing instances per IP to avoid hitting S3ās limit. Could you share any insights or recommendations on this?
from datatrove.
Thank you for sharing your experience!
from datatrove.
With the implementation of this feature and increasing the wait time, we've noticed a corresponding improvement in network speeds. It seems likely that there were existing bottlenecks due to excessive API calls. (wait time 180 -> 2500, network speed 180MBit/s -> 1.4GBit/s)
In light of this, it might be beneficial to include log messages that reflect these changes or provide guidelines for adequately setting the randomize_start_duration to prevent such bottlenecks.
And also I will do additional PR to examples/{some codes} for fixing randomize_start
parameter, this weekend.
from datatrove.
Related Issues (20)
- MinhashDedup should use the language parameter
- I would like to get help from Datatrove enthusiasts regarding issues I'm facing while running the example script. HOT 33
- SlurmExecutor error: `TypeError: cannot pickle 'module' object` HOT 3
- Batched filter inputs? HOT 5
- datatrove fails to handle tasks >1k with slurm job arrays HOT 25
- need for dependency check for using fasttext HOT 2
- LocalPipelineExecutor does not use cpu cores HOT 2
- how to postpone filter init till it's running HOT 5
- DataFolder.list_files fails when glob_pattern is specified and include_directories set to True
- Slurm Cluster Size for FineWeb
- adding an automatic log file `tail` under slurm executor HOT 1
- GopherQualityFilter.__init__() got an unexpected keyword argument 'language' on kaggle HOT 2
- solved: how to launch a slurm executor from an interactive slurm job
- Potential issue of dedup in index
- How about addding custom word_tokenizers?
- Add Dataset type parameters
- boto timeout when I read CC with Warc HOT 1
- no qos driving to invalid qos specification HOT 6
- Option to Not Upload to Hugging Face Hub in HuggingFaceDatasetWriter HOT 1
- Reading tokenized datasets HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
š Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. ššš
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ā¤ļø Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from datatrove.