Garbage Collection: Generational? Incremental? Both! — Mark Shannon’s proposal from Python Language Summit 2026

# What happened

Mark Shannon, the author of the incremental garbage collector that shipped in Python 3.14 and was later reverted, presented a path forward at the Python Language Summit 2026. His talk focused on reducing garbage-collection pause times while keeping memory use reasonable.

# Why it matters

# The trade-offs

Two mainstream GC strategies were compared:

• Generational GC assumes most objects die young, so it keeps memory use low but can produce long pauses when collecting older generations.

Shannon measured a metric he calls "scavenge effectiveness" (objects collected divided by objects visited). Both approaches had low effectiveness for Python: the generational GC was about 0.3% and the reverted incremental GC about 1%.

# Shannon's hybrid proposal

The recommended approach blends generational and incremental ideas:

• Use exactly two generations: a young and an old generation.
• Alternate between incremental scavenges on the young and old generations, avoiding full-heap non-generational scans.
• Fix the young generation initial size at 20 MB.
• Scavenge the old generation at twice the young-generation survivor rate (i.e., trigger old-gen work proportional to how many objects survive young-gen collections).

The goal is to reduce peak pause times like an incremental collector while keeping memory usage nearer to a generational collector.

# Questions raised and practical concerns

Na also asked about JVM-like runtime configuration for GC tuning. Shannon discussed the idea but noted that more configuration increases surface area and maintenance and that defaults should be improved first.

# Metrics and vocabulary to evaluate changes

Shannon emphasized the need for concrete metrics to judge new designs: pause time distributions, scavenge effectiveness, memory usage, and how collector behavior differs across real-world workloads. He introduced several helpful terms (scavenge, scavenge effectiveness, spaces, generations, survivor rates) to keep the discussion precise.

# Bottom line

Shannon's hybrid model aims to give Python lower pause times without the higher memory footprint of a fully non-generational incremental collector. The proposal is pragmatic: keep reference counting as the main collector, add a two-generation incremental scavenge schedule, and provide reasonable defaults (including a 20 MB young generation) while being mindful of maintenance cost and possible tuning knobs.