Tracking issue for only producing .xz on CI

[emilyalbini]

This issue tracks the work needed to only produce .xz tarballs on CI, recompressing them into .gz only during the release process. The effort will greatly reduce the amount of artifacts we need to store.

Implementation work needed:

• Change rust-installer to allow choosing which tarballs to produce Allow specifying the wanted compression formats at runtime rust-installer#105
• Switch CI to only produce .xz tarballs Refactor dist tarballs generation #79788 Only produce .xz tarballs on CI #80435
• Increase the specs of the VM running the release process to ensure recompression happens in a timely manner
• Retroactively delete .gz tarballs from the artifacts bucket

Bugs caused by this:

• Different contents of bare tarballs Use package name for top-level directory in bare tarballs #80397
• ./x.py install crashed due to a different directory structure x.py install fails when run outside of Rust source directory #80494 Fix broken ./x.py install #80514

[emilyalbini]

Forgot to create the tracking issue when the work started :)

[aidanhs]

I've got an inventory of the rust-lang-ci2 bucket. Observations:

• there are three 'categories' of gz
  1. rustc-builds-alt/.../*nightly*.gz - ~50k
  2. rustc-builds/.../*nightly*.gz - ~500k
  3. beta variants of i and ii - ~20k
  4. version numbered variants of i and ii - ~3k
• we're still uploading beta .gz files (but not release ones)

[aidanhs]

I think this is done. I will check the next inventory and, if so, revert rust-lang/simpleinfra@16e3aaa. For the record, here's my dirty script:

import boto3
import botocore
import csv
import json
import os
import collections
from concurrent.futures import ThreadPoolExecutor, as_completed

def mks3():
    session = boto3.Session(profile_name='mfa-rust')
    s3 = session.client('s3')
    return s3

status = json.load(open('progress.json'))
unsaved = 0
def save():
    global status
    print('Saving')
    json.dump(status, open('progress.tmp.json', 'w'))
    os.replace('progress.tmp.json', 'progress.json')
def set_deleted(key):
    global status
    global unsaved
    status[key] = True
    unsaved += 1
    if unsaved > 100:
        unsaved = 0
        save()

BUCKET = None
BATCH_SIZE = 100
NUM_WORKERS = 5
batches = []

def handle_batch(keys):
    global BUCKET
    s3 = mks3()
    to_delete = []
    deleted = []
    for key in keys:
        try:
            obj = s3.head_object(Bucket=BUCKET, Key=key)
            to_delete.append({ 'Key': key, 'VersionId': obj['VersionId'] })
        except botocore.exceptions.ClientError as e:
            if e.response['Error']['Code'] != '404':
                print('Failed to get object {} deets: {}'.format(key, e))
            else:
                print('Object {} missing, skipping'.format(key))
                deleted.append(key)
    if len(to_delete) != 0:
        print('Deleting {} objects'.format(len(to_delete)))
        ret = s3.delete_objects(Bucket=BUCKET, Delete={ 'Objects': to_delete })
        for obj in ret['Deleted']:
            deleted.append(obj['Key'])
        if 'Errors' in ret and len(ret['Errors']) != 0:
            print(ret['Errors'])
            assert False
    print('Finished deletion')
    return deleted

batch = []
rdr = csv.reader(open('nightlygzs'))
print('Creating batches')
for i, row in enumerate(rdr):
    if BUCKET is None:
        BUCKET = row[0]
    assert BUCKET == row[0]
    if row[1] in status:
        continue
    if len(batch) >= BATCH_SIZE:
        batches.append(batch)
        batch = []
    batch.append(row[1])

print('Got {} batches'.format(len(batches)))
executor = ThreadPoolExecutor(max_workers=NUM_WORKERS)
futures = [executor.submit(handle_batch, batch) for batch in batches]
print('Waiting for batch completion')
for i, complete in enumerate(as_completed(futures)):
    print('Completed batch {} of {}'.format(i+1, len(futures)))
    ret = complete.result()
    for key in ret:
        set_deleted(key)

save()

[aidanhs]

Correction - I deleted the nightly ones, so >500k combined. This looks like it's reduced our storage of ci2 from 78.5TiB to ~55TiB.

There are still a few other .gz files:

$ cat *.csv | grep '\.gz"$' | grep 'rustc-builds\(-alt\)*/.*/.*nightly.*\.gz"$' | wc -l
7090
$ cat *.csv | grep '\.gz"$' | grep -v 'rustc-builds\(-alt\)*/.*/.*nightly.*\.gz"$' | wc -l
26190

The first set are .gz files that mention nightly - not sure where they've come from, maybe they got missed in the initial run?
Of the second command, ~21k are nightly and most of the others are release versions. I estimate if we nuked the remainder of nightly/beta that's another 1TiB - but I'd wait until we stop uploading beta releases.

[jyn514]

What's the status of this? I think at this point we've stopped uploading new .gz files?

[emilyalbini]

This is done.