OFOffline First · Open source
Community homeHot discussions Latest posts Best of Knowledge base Resources About

Offline First · Open source · Why does offline-first-opensource use twice the memory of comparable tools?

1.0K
OFr/offline-first-opensource·posted by slow_query·3 days agoJobs

Why does offline-first-opensource use twice the memory of comparable tools?

Short version: offline-first-opensource needs almost no tuning at small and medium scale — the point where it starts to hurt is much further out than most people assume. Full measurements below.

One last trap: in container environments remember to adjust the memory-related parameters in step. Otherwise the host limit and the process expectation disagree, and the symptom is intermittent, unreproducible failure.

-- The query that broke: a full scan over 20M rows.
-- A composite index took P99 from 1.8s down to 42ms.
SELECT id, title, created_at
  FROM posts
 WHERE community_id = ?
   AND status = 1
 ORDER BY score DESC
 LIMIT 20;

We also fixed monitoring along the way: replaced average-based alerts with percentiles and split them per endpoint. False alerts dropped by about seventy percent and the on-call rotation visibly cheered up.

What genuinely surprised me was the tail. The average looked great while P99 jumped by an order of magnitude past some threshold. The cause was not offline-first-opensource itself but our upstream connection reuse — the load test traffic was too clean and hid the long-tail requests.

413 comments

413 comments

· first 120 loaded
M
Llinlin·3 minutes ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

517
Rran_bo·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

492
Kkite·2 days agoedited

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

488
Hhuang_ke·2 hours agoedited

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

466
Sswoole_lee·2 days ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

142
Mmike_xu·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

129
Cchen_devOP·2 days ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

43
Wwinter·2 days ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

223
Mmike_xu·2 days agoedited

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

448
Lli_ming·28 minutes agoedited

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

110
Nnikic·2 days agoedited

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

3
Ttang_hao·2 days ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

446
Lli_ming·just now

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

423
Mmike_xu·2 days ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

418
Aalice_dev·2 days ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

380
Sslow_query·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

155
Lli_ming·1 hour ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

367
Rran_bo·2 hours ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

350
Rran_bo·2 days ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

5
Nnikic·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

481
LlinlinOP·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

280
Rran_boOP·3 minutes ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

208
Ddev_zhou·2 days ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

137
Ttang_hao·12 minutes ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

264
Kkernel_panicOP·just now

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

192
Hhuang_ke·2 days agoedited

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

114
Llinlin·2 days ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

144
Zzhou_yi·2 days agoedited

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

344
Nnikic·28 minutes ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

251
Zzhu_zong·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

244
Zzhou_yi·2 days agoedited

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

236
Ddev_zhou·3 minutes ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

235
Bbob_chenOP·3 minutes ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

122
Rrase·2 days ago

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

146
KkiteOP·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

480
Wwinter·2 days ago

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

14
Rran_boOP·yesterday

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

169
Zzhu_zongOP·just now

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

413
RraseOP·5 hours ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

141
WwinterOP·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

426
Mmike_xu·2 days ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

1
Lli_mingMod·28 minutes agoedited

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

8
Hhuang_ke·2 days agoLevel 6

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

22
Zzhou_yi·2 days agoLevel 6

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

2
Zzhu_zongOPMod·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

5
Llinlin·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

1
Hhuang_ke·2 days agoedited

Saved. I am reworking this area this week — this saves a lot of wrong turns.

14
Ddev_zhou·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

3
Ttang_hao·2 days ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

213
Sswoole_lee·2 days ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

196
Sswoole_lee·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

171
Kkernel_panic·2 days agoedited

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

142
Ddev_zhou·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

111
Sslow_query·28 minutes ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

125
Ddev_zhou·2 days agoedited

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

420
Ddev_zhouMod·2 days agoedited

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

3
Mmike_xu·2 hours ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

148
Rran_bo·2 days ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

27
Cchen_dev·5 hours ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

96
Aalice_dev·2 hours ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

229
Sslow_query·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

167
Hhuang_ke·2 days agoLevel 6

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

141
Zzhou_yi·2 days ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

14
Rran_bo·just now

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

61
Kkernel_panic·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

6
Zzhou_yi·2 days agoLevel 6

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

238
Zzhou_yiOP·28 minutes agoeditedLevel 6

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

6
Ddev_zhou·3 minutes ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

48
Nnikic·2 days ago

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

278
Llinlin·5 hours ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

88
Zzhu_zong·2 days ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

28
Mmike_xuOP·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

34
Cchen_dev·2 days ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

3
Rrase·2 days ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

90
Hhuang_ke·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

286
Rrase·2 days agoLevel 6

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

14
Bbob_chenOP·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

5
Zzhou_yi·2 days ago

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

1
Sslow_queryMod·2 days agoedited

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

269
Ttang_hao·2 days agoLevel 6

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

391
Cchen_dev·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

11
Nnikic·2 days agoLevel 6

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

75
Rran_bo·12 minutes ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

116
Ttang_hao·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

42
Zzhu_zongOP·5 hours agoedited

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

55
Kkite·12 minutes agoedited

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

92
Ttang_hao·2 days ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

78
Zzhu_zong·2 days ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

70
Ttang_hao·2 days agoedited

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

4
Nnikic·2 days agoedited

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

4
Ddev_zhou·28 minutes ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

67
Nnikic·2 days ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

65
Kkite·2 days ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

64
Rran_boOP·2 days agoedited

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

2
Aalice_dev·12 minutes ago

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

60
Rran_bo·2 days ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

46
Llinlin·5 hours ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

42
Wwinter·2 days ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

32
Zzhu_zong·2 days ago

I see point 3 differently. The trade-off depends on your read/write ratio: read-heavy with little writing means caching actually widens the inconsistency window.

21
Mmike_xu·2 days ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

175
Nnikic·2 days ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

18
Nnikic·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

16
Rrase·28 minutes ago

Saved. I am reworking this area this week — this saves a lot of wrong turns.

6
Mmike_xu·3 minutes ago

Agreeing with the above. One addition: with this option enabled the GC count in your metrics doubles, so adjust the alert threshold at the same time or it will keep firing.

181
Cchen_devOP·2 days ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

198
Hhuang_ke·2 days ago

This matches what we see in production. We only hit it past 3k QPS; the earlier load tests showed nothing — the test traffic was too clean, with no long-tail requests.

1
Rran_bo·2 days ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

444
Ddev_zhou·2 days ago

There is actually a simpler fix that needs no architecture change: move this check up to the gateway and the problem disappears. The cost is one extra lookup at the gateway.

1
Cchen_dev·28 minutes agoedited

Has anyone run a controlled experiment? I did, reducing it to a single variable, and the difference was 4% — within noise. So I suspect the main cause is something else.

104
Zzhu_zong·28 minutes ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

350
Bbob_chen·2 days ago

Sharing our numbers, 8 cores 16GB, same scenario:

| Concurrency | P50 | P99 |
|---|---|---|
| 200 | 12ms | 88ms |
| 500 | 31ms | 340ms |

P99 clearly collapses at 500 concurrency, which lines up with your knee point.

361
Ddev_zhouOP·2 days ago

Worth learning from this debugging approach. We went straight at the logs and took a much longer route.

135
Mmike_xuMod·2 hours ago

A question: what changes in a container with a 512Mi memory limit? That is how we run it in production.

86
Lli_ming·2 hours ago

I just read the offline-first-opensource source — the author actually explains the reasoning in a comment, roughly "so that it degrades into predictable behaviour in extreme cases".

4
Cchen_dev·12 minutes ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

477
Nnikic·2 days agoeditedLevel 6

This is not a offline-first-opensource problem, it is a usage problem. The docs say this API is not thread-safe and you must lock around it yourself.

1
Lli_ming·2 hours ago

One counter-example: below offline-first-opensource 7.4 the semantics of that code are different, so do not copy it verbatim. We got burned in staging and rolled back once.

247
Llinlin·2 days ago

Can you give a minimal reproduction? I ran it locally for ten minutes and could not reproduce on macOS with the latest version.

6
Zzhu_zong·12 minutes ago

We have run this in production for two years without hitting it. That said, we never reached this scale, so our experience is not really evidence here.

1
Rrase·12 minutes ago

Thanks for sharing real numbers — far more useful than the articles that only cover concepts.

464

This is the post detail page /en/c/offline-first-opensource/post/p1. Posts and comments are generated deterministically from a seeded PRNG, so the same post always renders the same content and the link can be shared, reloaded and indexed. In production this page reads MySQL for the post, Redis for hot-post caching, and fetches the whole comment tree in a single query on the path column.

See the database schema →