The most-upvoted complaint in 18,464 habit-app reviews is "I can't fix yesterday" — so I built the fix

Chris Cheng — Thu, 12 Nov 2026 13:15:55 +0000

Every habit tracker in the top search results has a rating between 4.3 and 4.95 stars. If you're thinking about building one, that looks like a market with no room left.

Two developers recently did the work of actually reading the reviews instead of trusting the stars — and I want to add one thing they both surfaced but neither could ship a fix for yet.

The data, credited, not mine:

Matheus Coelho read the 100 most recent reviews for the top 4 "habit tracker" results on both stores — 8 apps, 800 reviews (link, published 2026-10-14). Of the 80 one- and two-star reviews, the top complaint was subscription / paywall / paid features (22), ahead of bugs (12) and sync or lost data (10). His example quotes: "paywalled at 3 habits", "must pay to track more than 4".
Eltacrew pulled every public review he could for six of the biggest habit trackers on Google Play — 18,464 reviews, 1,920 of them rated 1–3 stars (link, data collected 2026-09-12). Same winner by count: paywall / subscription, 450 reviews and 1,197 helpful votes. Notifications that don't fire came second (267 reviews).

So paywalls are the #1 complaint by volume. That's the headline, and it's true. But it isn't the most interesting line in either dataset.

The single most-upvoted review in 18,464 reviews isn't about money at all.

119 helpful votes, on Loop (the open-source one, 5M+ installs), translated from Korean:

"If I forget to check or check the wrong thing and the day changes, the previous day isn't shown, so I can't check it or fix it. Annoying."

Nothing else in 18,464 reviews got close to that vote count. It's not a crash, not a paywall — it's I missed a tap at 11:58 pm and now my record is wrong forever. Eltacrew's own takeaway, his #1 of four: "Let people fix yesterday. The most-agreed-with complaint in the category is a missing edit affordance."

That rings true to me as a builder, because a streak isn't a score. It's the app's memory of your life. When the memory is wrong and can't be corrected, people don't just lose a number — they stop trusting the record, and then they stop opening the app. The paywall complaints are mostly about the same thing, really: the limit appeared after I'd already invested days of data. Both complaints are about the app rewriting your history without asking.

So I built the fix — and its limit matters as much as the fix.

I'm the solo developer of DropDrop, a habit tracker for iOS. In version 1.2.4 (released 2026-10-31) I shipped Streak Protection & Repair: protect a streak before life happens, or repair it after — plus a quiet "rest" state for skipped habits, instead of a broken streak. A scheduled day you deliberately pause can be recorded as an intentional skip, with a reason: it protects the streak without adding a fake completed day, and it stays excluded from your completion target. The original schedule stays visible in the raw data.

The limit, stated plainly because it matters: a miss stays a miss unless you repair it. I deliberately did not build free backfill of invented completions — if every miss can be relabelled after the fact, completion rates stop telling you anything true. The repair exists for the 11:58 pm tap, not for rewriting a week you didn't do. (My longer write-up on rest days, skips and misses.)

Two things from the data I haven't solved, honestly:

Notifications. The second-biggest complaint in the 18,464-review set (267 reviews) is alarms that don't fire or fire wrong — and on Android that's partly the OS, not the app. On iOS I schedule each reminder individually, but I won't claim the problem is fully solved.
The paywall surprise. The anger isn't about price, it's about the limit appearing on day three. DropDrop's free tier is generous, but the test is whether a new user can see the limit on day one, before they've invested days of data. I'm still checking that onboarding flow against this standard.

What I'd take from the data if you're building or choosing a habit app:

Let people fix yesterday. It's the most-agreed-with complaint in the category.
If there's a limit, show it on day one, not day three.
A habit app's notification is its product.
The tap that marks a habit done must not wait on the network.

The data belongs to the two posts linked above — all I did was read them, take the finding seriously, and ship the one fix that was in my power to ship. If your app already lets people fix yesterday, tell me how you handle the honesty problem: where do you draw the line between repairing a missed tap and letting someone rewrite history?

DropDrop: https://www.dropdrophabit.com/ (App Store: search "DropDrop: HabitTracker & Todo")
I'm Chris Cheng, solo developer of DropDrop since June 2025. The 1.2.4 release notes above are from our Product Hunt launch; the review numbers in this post are from the two linked DEV posts, not my own scrape.

Deep Dive: Meta's AI Infrastructure and Developer Tools - A Technical Analysis

Chris Cheng — Fri, 27 Dec 2024 09:58:01 +0000

Introduction

As a developer working with AI infrastructure, I've been analyzing Meta's recent developments in artificial intelligence, particularly their open-source contributions and developer tools. This technical deep-dive explores the architecture, implementation details, and practical applications of Meta's AI ecosystem.

Technical Architecture Overview

Infrastructure Components

Example of Meta's distributed training configuration
config = {
'model_parallel_size': 8,
'pipeline_parallel_size': 4,
'num_gpus': 16000,
'optimizer': {
'type': 'AdamW',
'lr': 1e-4,
'weight_decay': 0.01,
'fsdp_config': {
'sharding_strategy': 'FULL_SHARD',
'mixed_precision': True
    }
  }
}

Key Technical Features

Distributed Training Infrastructure
- FSDP (Fully Sharded Data Parallel) implementation
- Custom memory optimization techniques
- Advanced pipeline parallelism
Model Architecture Innovations
- Grouped-query attention mechanisms
- Rotary position embeddings
- Custom normalization layers

Developer Tools and APIs

PyTorch Integration

from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy
def create_model():
model = MyLargeModel()
wrapped_model = FSDP(
model,
auto_wrap_policy=size_based_auto_wrap_policy,
mixed_precision=True
)
return wrapped_model

Performance Optimizations

Memory Efficiency:
- 40% reduction in memory usage
- Improved throughput with custom CUDA kernels
- Dynamic memory management
Training Speed:
- 2.5x faster training with optimized data loading
- Custom gradient accumulation
- Efficient parameter sharding

Practical Applications

Use Cases in Production

Large-scale model training
Real-time inference optimization
Multi-modal AI applications

Code Example: Optimized Attention Implementation

class OptimizedAttention(nn.Module):
def init(self, dim, num_heads=8):
super().init()
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim -0.5
self.qkv = nn.Linear(dim, dim 3, bias=False)
self.proj = nn.Linear(dim, dim)
def forward(self, x):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
q, k, v = qkv.unbind(2)
# Efficient attention computation
attn = (q @ k.transpose(-2, -1)) self.scale
attn = attn.softmax(dim=-1)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
return self.proj(x)

Best Practices and Guidelines

Infrastructure Setup
- GPU cluster configuration
- Network optimization
- Storage architecture
Model Development
- Code optimization techniques
- Memory management strategies
- Performance monitoring

Future Developments

Next-generation attention mechanisms
Advanced distributed training techniques
Improved developer tools and APIs

Conclusion

Meta's AI infrastructure represents a significant advancement in large-scale AI development. The combination of efficient architecture, optimized implementations, and developer-friendly tools makes it a powerful platform for AI practitioners.

DEV Community: Chris Cheng