## TL;DR
Long runs hit the per-series point cap when logging every step for many epochs: thin the series by logging every N steps, aggregate window statistics instead of raw points, or split the run per phase. Check the series length that failed against the documented maximum and redesign the logging cadence so the longest planned run stays under it.

```text
Neptune run stopped error 'InternalError: Series exceeds maximum number of points' failed
```

## Use this when
- The run or call fails with this symptom: InternalError: Series exceeds maximum number of points from a Neptune run.
- You are working on a Neptune run logging per-step metrics over a long training and the failure repeats on the same run or configuration.
- You need the failing run, call, or stage isolated before changing code or thresholds.

## Not for this skill when
- A series exploding because step values repeat due to a counter bug, where the fix is the step counter rather than thinning.
- The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

## Steps
1. Find the series named in the run and count its points; compare with the documented per-series maximum. Verify: the over-limit series and its count are identified
2. Compute the planned run's total points at the current cadence and confirm it exceeds the cap before training ends. Verify: the cadence math shows the overflow point
3. Change logging to every N steps or windowed aggregates so the longest run lands below the cap with headroom. Verify: projected points sit under the limit
4. Restart or continue logging with the new cadence and confirm points accumulate without the error. Verify: the series grows cleanly past the old failure point

## Variant phrasings
### neptune series exceeds maximum number of points
The verbatim phrasing.
### neptune run stopped too many points per step logging
The cadence phrasing.
### neptune series limit long training run fix
The long-run phrasing.

Compatibility: Neptune long-run series limits

## Why it happens
Series storage is provisioned per series with a hard point ceiling. Per-step logging feels free early in a run, so the ceiling arrives mid-training as a surprise, exactly when the accumulated history is most valuable.

## Edge cases / pitfalls
- Evaluation points logged per batch are the usual hidden multiplier; epochs, not steps, are the right cadence for them.
- Splitting a run loses one continuous history; prefer thinning unless phases are genuinely separate.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_haQkuYYjyUYh7luHrkEGRw