Very long runtime (>20 hours) for feature-classifier extract-reads with SILVA 138.2 in QIIME 2 2026.4

Technical Support
1

Hello,

I am experiencing an unusually long runtime when running qiime feature-classifier extract-reads with SILVA 138.2 on QIIME 2 2026.4.

My command is:

qiime feature-classifier extract-reads \
  --i-sequences silva-138.2-dna-seqs.qza \
  --p-f-primer CCTACGGGNGGCWGCAG \
  --p-r-primer GACTACHVGGGTATCTAATCC \
  --p-max-length 600 \
  --o-reads silva-138.2-99-V3V4.qza

Environment

  • QIIME 2: 2026.4
  • Python: 3.12.13
  • OS: Ubuntu on WSL2
  • CPU: Intel Core i9
  • RAM: 80 GB

Input data

The reference sequences were generated using RESCRIPt:

qiime rescript reverse-transcribe \
  --i-rna-sequences silva-138.2-seqs.qza \
  --o-dna-sequences silva-138.2-dna-seqs.qza

The input artifact contains approximately 510,495 sequences.

What I observed

  • Running the same command on a test dataset containing only 100 sequences finishes within a few seconds.
  • On the full dataset, the process continues for more than 20 hours.
  • During execution:
    • CPU usage stays around 100–103%
    • Memory usage remains stable at about 1.2 GB
    • I/O wait is 0%
    • No output artifact (silva-138.2-99-V3V4.qza) is created while the process is running.
  • Using --verbose does not produce any progress messages.

Current process status:

ELAPSED   21:53:02
%CPU      102
TIME      22:26:53
RSS       1244088 KB

Disk space and memory are sufficient.

Question

Is this runtime expected for approximately 510k SILVA sequences, or could this indicate a performance issue in feature-classifier extract-reads in QIIME 2 2026.4?

Has anyone observed similar behavior?

Thank you for any suggestions.

2

Additional information:
I tested the exact same input artifact and command on a Mac (QIIME 2 2026.4), and feature-classifier extract-reads completed successfully in a reasonable amount of time. The issue only occurs on my Ubuntu WSL2 environment, where the process runs for over 20 hours with ~100% CPU utilization and no output produced.

3

If you want to post the artifacts to the forum publicly, we can take a look at the history and see what could have caused this.

I also have run Qiime2 using WSL2 and using a Mac, and I find the mac to be much faster!

Oh, there's a function of rescript that does this with vsearch and it's much faster! It's even multi-threaded, so make sure to use that too!

Try that and report back:

4

Hi Colin,

Thank you for your suggestion.

I prepared a zip file containing the input artifacts and environment information:

  • command.txt
  • conda-list.txt
  • cpu-info.txt
  • md5sum.txt
  • qiime-info.txt
  • silva-138.2-dna-seqs.qza
  • silva-138.2-seqs.qza

The file is available here:

The download link will be available for one week.

Could you please take a look at the artifacts and let me know if you can identify what might be causing the issue on my WSL2 environment?

Thank you very much for your help.

5

Thank you for this detailed information!

I was able to get this working on my laptop. Full writeup on GitHub:

Remarkably, the vsearch accelerated rescript extract-seq-segments takes just as long as the built-in feature-classifier extract-reads feature, at about 5 minutes each.

EDIT: I'm not sure I really answered your question as this also works on your mac. I don't have access to my WSL2 machine, right now! I'll have to check later or wait for someone else to test this on WSL2

6

Hi Colin,

I tried the rescript extract-seq-segments approach, and it worked perfectly.

In the end, I wasn't able to get feature-classifier extract-reads to work, so I'll use the RESCRIPt approach instead.

Thanks so much for your helpful suggestion!