Skip to content
Permalink

Comparing changes

Choose two branches to see what’s changed or to start a new pull request. If you need to, you can also or learn more about diff comparisons.

Open a pull request

Create a new pull request by comparing changes across two branches. If you need to, you can also . Learn more about diff comparisons here.
base repository: ipfs/someguy
Failed to load repositories. Confirm that selected base ref is valid, then try again.
Loading
base: v0.14.1
Choose a base ref
...
head repository: ipfs/someguy
Failed to load repositories. Confirm that selected head ref is valid, then try again.
Loading
compare: v0.15.0
Choose a head ref
  • 3 commits
  • 15 files changed
  • 1 contributor

Commits on Jul 27, 2026

  1. fix: stream ndjson and cap background lookups (#169)

    * fix: stream ndjson instead of buffering it
    
    Response compression withheld writes until 200 bytes had accumulated
    before deciding whether to compress, and Flush is a no-op until that
    decision is made. Provider records are usually smaller than that, so a
    provider someguy had already resolved sat in the buffer while it looked
    up the rest, and the response headers were withheld too. Clients that
    bound the whole request, such as Helia's delegated routing client at
    30s, gave up and saw an empty result even though providers were found.
    
    - compress from the first byte so the per-record flush takes effect
    - stop routing lookups at 25s, leaving room to flush before clients
      that allow 30s walk away
    
    * feat: cap concurrent background peer lookups
    
    Provider records that arrive without addresses trigger a FindPeer that
    runs detached from the request, so a client can close its connection and
    leave a full DHT walk behind. Nothing bounded how many ran at once: a
    cheap request in, an unbounded pile of DHT walks out, on hosts that
    already sit near their memory limit.
    
    Cap them at 512 per instance and skip the lookup when full, the same
    outcome as a peer under connect-failure backoff. Production dispatches
    roughly 0.3-0.5 lookups per second per instance, which is about 20
    concurrent even if every one runs to its timeout, so the cap only
    engages far outside normal traffic.
    
    - reserve the slot before counting the lookup, so a skipped dispatch
      cannot stall the iterator waiting on a lookup that never ran
    - add in-flight, rejected, and duration metrics to size the cap
    
    * docs: document background peer lookup metrics
    
    Cover the three metrics added with the lookup cap, and note the cap in
    the caching doc next to the dispatch it bounds.
    
    * feat: make lookup cap and routing timeout configurable
    
    Both were compile-time constants, so an operator hitting the lookup cap
    had to rebuild and redeploy to raise it. That defeats the point of the
    cap: it exists as a safety valve, and the metrics that show it engaging
    are only useful if the limit can be changed in response.
    
    - SOMEGUY_CACHED_ADDR_BOOK_MAX_CONCURRENT_FIND_PEERS, default 512
    - SOMEGUY_ROUTING_TIMEOUT, default 25s, treats 0 as unset rather than
      cancelling every request before it reaches a router
    lidel authored Jul 27, 2026
    Configuration menu
    Copy the full SHA
    d18d0b1 View commit details
    Browse the repository at this point in the history
  2. chore: bump boxo to main and go-libp2p-kad-dht v0.42.1 (#171)

    Picks up the shorter /routing/v1 stale windows from ipfs/boxo#1195, so
    caches stop serving a two-day-old answer while someguy is healthy.
    
    go-libp2p-kad-dht v0.42.0 dropped the context argument from its
    constructors. Cancelling the context that built a DHT no longer stops
    it, so someguy lost its only shutdown path and had to grow an explicit
    one, the same way kubo closes both clients on stop.
    
    - drop ctx from dht.New, and from newBundledDHT which no longer used it
    - add bundledDHT.Close and call it on shutdown, covering the standard
      client too via io.Closer
    - update the Cache-Control assertions the new stale windows changed
    - restore the changelog section order broken in #169
    lidel authored Jul 27, 2026
    Configuration menu
    Copy the full SHA
    248a3e6 View commit details
    Browse the repository at this point in the history
  3. chore: release v0.15.0 (#172)

    Move the accumulated Unreleased entries into a dated section, the same
    way v0.14.0 was cut.
    
    The relay address TTL and address ordering entries came from #159, which
    shipped in v0.14.1. That release only bumped version.json, so it never
    got a changelog section and those two sat under Unreleased. They are
    filed under v0.14.1 rather than v0.15.0, so anyone already running
    v0.14.1 is not told they are new.
    lidel authored Jul 27, 2026
    Configuration menu
    Copy the full SHA
    8585f9b View commit details
    Browse the repository at this point in the history
Loading
⌂ 换网址