Nano 1.3 Batched Sync Flush: Memory and Latency Trade-offs

1. Expected Memory Overhead

In Nano 1.3 each queued entry is kept in memory until the synchronous flush commits it. A rough per‑entry footprint of 64 bytes (object header + references + serialized payload) yields:

Batch Size	Approx. Memory
1 000 entries	≈64 kB
5 000 entries	≈320 kB

Thus moving from 1 000 to 5 000 entries multiplies the peak memory usage by roughly five. On devices with limited heap this can trigger OOM if the application is already near capacity.

2. Latency Across Entry Sizes

The batched sync flush writes the whole batch in a single I/O transaction. For small‑to‑medium payloads the per‑entry latency is dominated by the fixed I/O latency, so it stays relatively flat as entry size grows. When entries become large (several kilobytes), serialization and buffer copying add a modest overhead, raising per‑entry latency by a few microseconds.

Because the number of round‑trips remains one, the total flush time grows sub‑linearly: doubling the batch size from 1 000 to 5 000 typically adds ~10–20 % to the overall flush duration, not five times.

Verification Steps

1. Profile heap usage while queuing 1 000 and 5 000 entries. Tools: VisualVM, YourKit, or Android Studio profiler. Compare peak memory delta to the 5× estimate.

2. Measure flush latency with a high‑resolution timer. Example (Java):

   long start = System.nanoTime(); nano.flush(); long elapsed = System.nanoTime() - start;

   Run for batches of 1 000, 2 000, 5 000 entries and plot per‑entry latency against entry size.

3. Use a disk‑I/O monitor (iostat, perf) to confirm that the flush performs a single write operation regardless of batch size.

4. On Android or embedded Java, watch GC activity during flushes to see if the larger batch triggers additional collection pauses.

Missing Diagnostic Detail

To fine‑tune the recommendation, could you share the average serialized size of your entries (in bytes)? This will help estimate the exact memory footprint and identify whether the 64‑byte assumption holds.